Facial approximation (FA) is a common tool used to recreate the possible facial appearance of a deceased person based on the relationship between soft tissue and the skull. Although this technique has been primarily applied to modern humans in the realm of forensic science and archaeology, only a few studies have attempted to produce FAs for archaic humans. This study presented a computerized FA approach for archaic humans based on the assumption that the facial soft tissue thick- ness depths (FSTDs) of modern living humans are similar to those of archaic humans. Additionally, we employed geometric morphometrics (GM) to examine the geometric morphological variations between the approximated faces and modern human faces. Our method has been applied to the Jinniushan (JNS) 1 archaic human, which is one of the most important fossils of the Middle Pleistocene, dating back to approximately 260,000 BP. The overall shape of the approximated face has a relatively lower forehead and robust eyebrows; a protruding, wider, and elongated middle and upper face; and a broad and short nose. Results also indicate skull morphology and the distribution of FSTDs influence the approximated face. These experiments demonstrate that the proposed method can approximate a plausible and reproducible face of an archaic human. Keywords Facial approximation · Archaic humans · Facial soft tissue thickness depths · Geometric correspondences · Assessment Introduction Facial approximation (FA) or craniofacial reconstruction aims at recreating a potential facial appearance from a dry skull. This technique is often the last hope in the realm of forensic science when no other clues and evidence support * Wuyang Shui the investigation and identification (Wilkinson 2010). Based sissun@126.com on the assumed relationship between soft tissue and the bony structure, FA has been applied in archaeology to reconstruct Department of Archaeology, University of York, The King’s the portraits and facial appearances of people in the past Manor, York YO1 7EP, UK (Kustar 2004; Benazzi et al. 2009; Hayes, et al. 2017; Shui Key Laboratory of Vertebrate Evolution and Human Origins and Wu 2018; Marić et al. 2020). It has sometimes been of Chinese Academy of Sciences, Institute of Vertebrate Paleontology and Paleoanthropology, Chinese Academy applied to named individuals, but more usually unnamed of Sciences, Beijing 100044, China people from the past. Nonetheless, these applications always Joint International Research Laboratory of Environmental focused on modern humans, and they are less commonly and Social Archaeology, Shandong University, applied to archaic humans, where differences in skull and Qingdao 266237, Shandong, China facial morphologies make the approximation more challeng- Institute of Cultural Heritage, Shandong University, ing. In recent years, the approximated appearance of our fos- Qingdao 266237, Shandong, China sil ancestor has become an area of study for anthropologists CAS Center for Excellence in Life and Paleoenvironment, and has also captured the imagination of the general public, Beijing 100044, China influencing perceptions of how “like us” and how “human” College of Information Science and Technology, Northwest Neanderthals were. The visualization of the approximated University, Xi’an 710127, China Vol.:(0123456789) 1 3 186 Page 2 of 20 Archaeol Anthropol Sci (2021) 13:186 face rather than an imaginary approximation provides an Pioneering work on 3D graphical computerized FA was effective 3D presentation to help us perceive and understand first proposed by Vanezis ( 1989). The average FSTDs at a the characteristic features of human fossils. In addition, FA limited number of anatomical landmarks were used to pro- offers a new insight to investigate the morphological shape duce a coarse mask, and then the generic face was deformed variations between archaic humans and Homo sapiens. to recreate a facial appearance over the dry skull. It is 3D manual facial approximation approaches have been acknowledged that the greater number of FSTDs is acquired, widely used to recreate facial appearances (Hayes 2016). the greater reliability of the approximation is achieved. Anthropologists collaborated with artists to recreate a pos- Another effective computerized FA employed the deforma- sible likeness by means of modeling clay or plasticine over tion-based approach based on the assumption the verified the replica of the skull and adding the facial features, e.g. craniofacial relationship of the template model is similar to eyes, nose, and mouth. During this procedure, muscle struc- that of the dry skull, removing the skull morphology varia- tures and facial soft tissue thickness depths (FSTDs) at ana- tions (Quatrehomme et al. 1997; Nelson and Michael 1998; tomical landmarks can be used to represent the craniofacial Turner et al. 2005; Deng et al. 2011; De Buhan and Nardoni relationship between soft tissue and skull. The manual FA 2018). In this procedure, either a generic face or a specific approaches can be divided into three main categories: the face based on the properties of the dry skull, e.g. age, sex, Russian anatomical approach, the American anthropomet- and ethnic group, was chosen as the template model. Then, rical approach, and the combination Manchester approach the template face was deformed following the same transfor- (Verzé 2009). However, they are heavily dependent on the mation that was calculated by deforming the template skull degree of anthropological interpretation and the practition- to the dry skull. This approach is simple and easy-to-use, ers’ subjective experience. Under such circumstances, mul- because it does not require the FSTDs table at anatomical tiple approximated faces of the same skull can be produced. landmarks. In recent years, with the increasing availability For instance, three portraits of Ferrante Gonzaga, an Italian of skull and face datasets of modern living humans, a regres- nobleman of the Renaissance, have been recreated (Fatuzzo sion-based method has been applied to study the craniofacial et al. 2016). Such various approximations with inconsist- relationship based on principal components (PC) scores of ent facial features might probably lead to less public con- every skull and face in the shape space (Paysan et al.2009; fidence when no convincing hypothesis, and evidence can Berar et al, 2011; Deng et al. 2016). Then, this predicted be provided. craniofacial relationship can be used to recreate the facial With the rapid progress in computer science and medical appearance. image acquisition, computerized FA technology has been With regard to the similarities in the craniofacial rela- gradually developed to increase the level of accuracy and tionship between relatively recent modern human remains reliability of the approximated face. The basic idea is to and modern living humans, the majority of existing studies mimic the manual FA approach using the computer (Wilkin- focused mainly on approximating the appearance of archaeo- son 2005). Using the FSTDs at anatomical landmarks and logical human fossils. In contrast, only a few publications knowledge of facial muscle, both 2D and 3D interactive concerned the investigation of archaic humans (Hayes 2016). graphics technologies which mimic the Manchester approach Hayes et al. estimated the frontal and lateral appearances of have been developed. In 2D interactive FA, the frontal and Liang Bua, the holotype of Homo floresiensis (Hayes et al, profile portraits were recreated using Adobe Photoshop soft- 2013). In their work, the 2D profile outlines of the approxi- ware (Hayes et al. 2012). Likewise, 3D interactive FA was mated face were created based on the FSTDs at landmarks. used to recreate a 3D probable likeness through a haptic Then, muscle images were deformed and attached to the sur- feedback device and 3D software, e.g. Autodesk 3ds Max, face of the skull. Finally, the reliability of the reconstructed ZBrush, and Blender (Wilkinson et al. 2006; Lee et al. 2014; face was evaluated using geometric morphometrics (GM). Short et al. 2014; Miranda et al. 2018). In their work, the Because 3D facial morphology might allow anthropolo- tissue depth pegs which represented the FSTDs at anatomi- gists to better elucidate the facial characteristics of archaic cal landmarks were attached to the correspondence vertices humans and investigate evolutionary changes in the face, of the dry skull, and the facial muscles were revised and a 3D computerized FA approach still needs to be further attached to the surface of the skull. Then, the facial features investigated. were added and sculpted to improve the reliability of the The Jinniushan (JNS) 1 cranium, dating back to approximated face. However, all these technologies require 260,000 years BP, was discovered in Yingkou County, both anatomical knowledge and expertise in modeling skills. Liaoning Province, in northeast China in 1984 (Wu 1988; Anthropologists have to invest great effort in manual mode- Rosenberg et al. 2006). It is one of the most important fossils ling when they wish to produce a range of multiple candidate in East Asia and has been used to investigate morphologi- faces that use the FSTDs of different samples. cal features and shape variations with other fossils (Hublin 2013; Athreya and Wu 2017). It appears that its supraorbital 1 3 Archaeol Anthropol Sci (2021) 13:186 Page 3 of 20 186 shape, superciliary arch thickness and shape, postorbital The Tabun 2 mandible was found in stratigraphic layer constriction, and paranasal inflation are somewhat closer C of the Tabun cave, one of the paleoanthropological sites to those of Dali and Maba individuals, who are considered in the Near East. It was reconstructed and virtually recov- to represent population immigration from outside of China, ered in six fragments, but lacking the left condyle, part of and to be the result of an admixture with archaic humans basilar symphysis (Schwartz and Tattersall 2000, Quam and (Andrews, 1986; Rightmire 1998). Although a manual Smith, 1998). Morphologically, Tabun 2 is relatively large approach has been used to produce the facial appearance of and robust, and it indicates a strong development of ante- JNS 1, considerable interest has been shown in investigat- rior marginal tubercle and a triangular basal corpus profile ing the approximated face of JNS 1 based on reasonable at the symphysis and mandibular foramina. It exhibited a assumption and supporting data, rather than experience and mixture of morphological features of Neanderthals and early imagination. This paper aims to provide a computerized FA modern humans (Harvati and Lopez 2017). In addition, the method to approximate the plausible and reproducible face well-preserved and complete Mauer 1 mandible (Wagner of the archaic human. et al. 2010), a holotype of Homo heidelbergensis, was found near Mauer, southeast of Heidelberg, Germany in 1907. It is the oldest hominin fossil reported to date from central and Materials and methods northern Europe. It is of note that Mauer 1 exhibits a mix- ture of both primitive and modern features (Mounier et al. Materials 2009). In this study, Tabun 2 and Mauer 1 were selected to fit with the JNS 1 cranium. Then, these two reassembled The archaic human fossil skulls (called JNS 1 using Tabun 2 and JNS 1 using Mauer 1) were used to approximate the face of JNS 1. Figure 1a The JNS 1 cranium retained most of the maxillary denti- shows the JNS 1 cranium (peach color) and the Tabun 2 tion although the bone has been broken into more than one mandible (gray color). Figure 1b displays the JNS 1 cranium hundred pieces (Wu 1988). In an attempt to perform FA suc- (peach color) and the Mauer 1 mandible (gray color). cessfully, the cranium required careful examination and res- toration. It has been manually repaired by researchers from Skull and face datasets of modern living humans the Institute of Vertebrate Paleontology and Paleoanthropol- ogy (IVPP) in Beijing, China. The restoration procedures In order to obtain the craniofacial relationship between soft were as follows: firstly, every fragment of JNS 1 fossil was tissue and skull, a total of 60 modern Chinese living humans cleaned and strengthened. Secondly, the fractured fragments (30 females and 30 males aged 20–30 years old), who lived were carefully matched together based on the similarity of in Shaanxi province in northern China, were enrolled in this the boundary of every fragment following the experience study. More details can be seen in our previous studies (Shui of the researchers. Thirdly, super glue was used to adhere et al. 2017). Each individual had normal morphology and fragments to each other. Finally, plaster was used to fill in had never undergone any orthodontic treatment. Medical the missing region of the cranium guided by geometric con- images were acquired by means of a clinical multi-slice CT straints. Anthropologists predicted the sex and age of JNS scanner system (Siemens Sensation 16). The CT images of 1 through the analysis of morphological features, sutures, each individual were archived in standard DICOM 3.0 with a and dental wear. In recent years, JNS 1 was suggested to resolution of 512 × 512 . All participants were provided with be female because of two important features, the subpubic full details of the study and written informed consent. This concavity and the medial aspect of the ischiopubic ramus research was approved by the Ethics Review Committee of (Rosenberg et al. 2006). Likewise, based on the comparison Department of Archaeology, University of York. of dentition and the analysis of tooth wear, an early study Our previous studies constructed dense corresponding suggested that JNS 1 was over 30 years old (Wu 1988), but vertices among skulls and faces, respectively. The pro- more recently, it was suggested to have been approximately cedure was as follows: firstly, image segmentation and about 20–30 years old (Herrera and Garcia-Bertrand 2018). the well-known marching cubes algorithm (Lorensen and Because only the JNS 1 cranium remained and the mandi- Cline 1987) were used to convert a series of CT images ble was not preserved, a well-preserved late archaic human to the digital skull (or face). Secondly, the external sur- mandible was required to assemble the JNS 1 cranium. But face of every skull (or face) within our dataset was com- it remains challenging to find a suitable mandible with simi- puted. Thirdly, anatomical landmarks of the skulls and lar age and features. We have to decide to use two archaic faces were defined and placed. Fourthly, iterative closest human mandibles whose ages covered the age of JNS 1, i.e. point (ICP), thin-plate splines (TPS), and compact support one mandible is more recent than JNS 1, and the other is radial basis function (CSRBF) algorithms were applied to older than JNS 1 to repair JNS 1. register all skulls (or faces). Next, the closest points were 1 3 186 Page 4 of 20 Archaeol Anthropol Sci (2021) 13:186 Fig. 1 JNS 1 cranium (peach color) and two different mandi- bles (gray color). (a) Tabun 2. (b) Mauer 1 considered as corresponding vertices, i.e. every skull (or configuration was defined. A total of 91 anatomical land- face), had the same number of vertices and each vertex of marks were chosen, and their 3D coordinates were acquired every skull (or face) was located approximately in cor- using Landmark Editor software (Wiley et al. 2005), where responding positions. To remove the effects of location, 17 anatomical landmarks were located on the midline and 74 orientation, and scaling, generalized Procrustes analysis anatomical landmarks were bilateral, respectively (Table 1). (GPA) and principal component analysis (PCA) were car- Most of these anatomical landmarks were defined accord- ried out to construct the skull and face statistical shape ing to Martin’s definitions (Martin 1928). Then, 404 semi- model. Every skull (or face) can be represented by the landmarks were placed on JNS 1, which were identified in coordinates of the average skull (or face) and the linear 16 patches based on the given landmarks. Here the semi- combinations of PC scores and corresponding orthogonal landmarks of each patch were equally spaced within a 3 × 3 PCs (Shui et al. 2020). frame as a patch, and each patch with less than 9 anatomi- cal landmarks was replenished with the middle points of Anatomical landmark definitions two adjacent anatomical landmarks (Table 2). Finally, these semi-landmarks were projected to the modern human skull, In order to estimate the overall shape of the facial appear- i.e. we established geometric correspondences between two ance, an anatomical landmark and semi-landmark skulls (Gunz and Mitteroecker 2013). 1 3 Archaeol Anthropol Sci (2021) 13:186 Page 5 of 20 186 Table 1 Anatomical landmarks No Landmark definition Position 1 Nasion Midline 2 Rhinion Midline 3 Nasospinale Midline 4 Subspinale Midline 5 Prosthion Midline 6 Infradentale anterius Midline 7 Gnathion Midline 8 Pogonion Midline 9 Glabella Midline 10 Ophryon Midline 11 Metopion Midline 12 Bregma Midline 13 Vertex Midline 14 Lambda Midline 15 Inion Midline 16 Intersection between inferior nuchal line and external occipital crest* Midline 17 Opisthion Midline 18 Maxillofrontale Bilateral 19 Orbitale Bilateral 20 Dacryon Bilateral 21 Superior orbital fissure (foramen), the closest point superior to the superior fissure* Bilateral 22 Most prominent point of supercillary above superior orbital fissure* Bilateral 23 Root of supercillary above superior orbital fissure* Bilateral 24 Ectoconchion Bilateral 25 Frontomalare orbitale Bilateral 26 Frontomalare temporale Bilateral 27 Zygion Bilateral 28 Zygomaxillare Bilateral 29 Jugale Bilateral 30 Most concave point on the inferior margin of maxilla* Bilateral 31 Vertical projection from jugale to lower margin of zygomatic* Bilateral 32 Infraorbital foramen Bilateral 33 Middle pyriform point, horizontal projection from infraorbital foramen to pyriform aperture* Bilateral 34 Lower pyriform point, lowest point of pyriform aperture* Bilateral 35 Stephanion Bilateral 36 Frontotemporale Bilateral 37 Auriculare Bilateral 38 Most prominent point on supramastoid crest* Bilateral 39 Asterion Bilateral 40 Mastoidale Bilateral 41 Ectomalare Bilateral 42 Coronion Bilateral 43 Lowest point of mandibular notch* Bilateral 44 Gonion Bilateral 45 Condylion laterale Bilateral 46 Ramus posterius, most concave point on the posterior margin of ramus* Bilateral 47 Ramus anterius, most concave point on the anterior margin of ramus* Bilateral 48 Vertical projection from lowest point of mandibular notch to lower margin of mandible along ramus Bilateral 49 Vertical projection from alveolare of lower m2 to lower margin of mandible* Bilateral 50 Temporale anterius, most anterior point of temporal squama* Bilateral 1 3 186 Page 6 of 20 Archaeol Anthropol Sci (2021) 13:186 Table 1 (continued) No Landmark definition Position 51 Temporale superius, most superior point of temporal squama* Bilateral 52 Alveolare of upper P3* Bilateral 53 Alveolare of lower m1* Bilateral 54 Mental foramen Bilateral 55 Mental laterale, turning point from mental to mandibular body on the inferior margin* Bilateral Anatomical landmarks are defined by the authors and the rest of anatomical landmarks without special notice are from Martin (1928) the template were calculated and visualized in a graphical Methods format; (c) the coarsely approximated face of JNS 1 was recreated by assigning dense FSTDs to the corresponding Figure 2 summarizes the framework of the proposed FA vertices of JNS 1. The TPS interpolation function was method. Firstly, the JNS 1 cranium and the selected man- used to improve the approximation. Due to the absence of dible were virtually reassembled, and the missing geom- the actual mandible, multiple approximations can also be etry was repaired. Secondly, a coarse-to-fine computer- mathematically calculated by interpolating the surfaces ized FA approach was proposed to recreate the possible of the approximated faces based upon Tabun 2 and Mauer likeness of JNS 1 based on the assumption that the distri- 1. (d) Quantitative evaluation was used to validate the bution of average FSTDs of the modern humans within reliability of the approximation through comparison of the dataset is similar to that of JNS 1. This procedure the distributions of FSTDs. Finally, we employed GM to comprised four steps: (a) a hybrid non-rigid registration examine the morphological shape variations between the approach was carried out to establish the dense geometric approximated faces and modern human faces. We exam- correspondences between the template skull and JNS 1, ined the effects of skull morphology and FSTDs on the where 495 landmarks and semi-landmarks were used to approximated faces. All these methods were programed guide the transformation mapping; (b) the dense FSTDs of Table 2 Eight patches of one No Patch of anatomical region Semi-landmark Numbers of anatomical land- side of JNS 1 density marks used in the patch 1 Zygomatic 5 × 5 19, 24, 25, 28, 29, 31 2 Maxilla 5 × 5 5, 28, 30, 32, 33, 34, 41, 51 3 Mandible 5 × 5 42, 43, 44, 45, 46, 47, 48, 49 4 Mental protuberance 5 × 4 6, 7, 8, 53, 54, 55 5 Nasal 5 × 3 1, 2, 18, 20, 33 6 Superciliary arch 5 × 5 9, 10, 18, 21, 22, 23, 25, 26, 36 7 Frontal 9 × 5 10, 11, 12, 35, 36 8 Parietal and occipital 11 × 6 12, 14, 15, 39, 50, 51 Fig. 2 The pipeline of the computerized facial approximation method 1 3 Archaeol Anthropol Sci (2021) 13:186 Page 7 of 20 186 using C + + and Matlab 2019. Taking the approximation overall shape of the face can be approximated based on skull of JNS 1 using Tabun 2 as an example, we introduce the morphology and the craniofacial relationship between soft proposed method. tissue and skull (Wilkinson 2005). We proposed a coarse- to-fine FA approach to produce a reproducible and objective approximation. The restoration of JNS 1 Geometric correspondences between the template skull Because the temporal mandible joint (TMJ) that con- and JNS 1 We employed two steps to establish high-qual- nected the JNS 1 cranium and Tabun 2 was insufficiently ity geometric correspondences between JNS 1 and the accurate, and the left condyle of Tabun 2 seemed incom- template skull. We assumed that all the landmarks and plete, the first step was to predict the missing geometry semi-landmarks of the template skull and JNS 1 were rep- of Tabun 2 and match the cranium and mandible closely. resented by = p , p ,… , p , p = x , y , z and 1 2 l i i i i Before the virtual restoration, the external surface of JNS = q , q ,… , q , q = x , y , z , respectively, where 1 2 l i i i i 1 was extracted. This comprised three steps: firstly, JNS l denoted the number of anatomical landmarks and semi- 1 was transformed into the Frankfort coordinate system landmarks. A popular non-rigid registration TPS function based on the left porion, right porion, left orbitale, and was first conducted to deform the template skull to JNS 1. the glabella. Secondly, an external point cloud was gener- During the deformation, it enabled the bending energy of the ated based on the cylindrical sampling algorithm. In this function f p = q minimized (Bookstein 1989). TPS can i i procedure, a couple of cross-section planes were gener- be represented by affine transformation parameters and non- ated between the bottom and top of JNS 1. For every affine warping parameters as the following linear equation: cross-section plane, the external points were obtained + ∙ by calculating the intersection points between JNS 1 and (1) a set of given rays, starting at centroid of every cross section along equally spaced angle vectors. Finally, the where the radial basis kernel can be represented by external point clouds were converted to a set of triangu- = K , K = ‖p − p ‖ , and ‖ ∙ ‖ denoted the Euclidean i,j i,j i j lar meshes. Subsequently, we employed the mirror resto- distance. was the regularization parameter that used to ration method to repair the external surface of Tabun 2 balance the smoothness. I denoted the l × l identity matrix; (Gunz et al. 2009). Figure 3 shows the external surface O denoted the 4 × 4 zero matrix; A denoted the 4 × 1 zero of JNS 1 using Tabun 2 that comprised the anatomical T T matrix; α = [a a a a ] ; and w = [ϖ ] represented the aff- 0 1 2 3 i landmark and semi-landmark configuration. ine and non-affine parameters, respectively. Following the same transformation that was computed Computerized facial approximation by warping the template skull to JNS 1, the template face was deformed to produce a possible likeness as a candi- In anthropology, it is widely accepted that facial surface date face. Next, we employed non-rigid registration to allow has a close relationship to the bony structure and that the Fig. 3 The restoration of JNS 1 using Tabun 2 that comprised the anatomical landmark and semi-landmark configuration 1 3 186 Page 8 of 20 Archaeol Anthropol Sci (2021) 13:186 the deformed template skull and JNS 1 to match closely a vertex of the template skull along the normal vector often by assigning an affine transformation to every vertex of the passed through the template face; thus, the intersection point deformed template skull. Assumed affine transformations can be calculated. The FSTDs were defined as the Euclidean X = [X X X …X ] of all the vertices, we defined the cost distances between pairs of corresponding vertices. It is of 1 2 3 n function, E(X), which consists of anatomical landmarks note that the normal vector of a given point that was deter- and semi-landmarks term E (X), a local affine regulariza- mined by the geometric coordinates and topologies of the tion E (X), and a stiffness term E (X). To evaluate the accu- neighboring vertices influenced the accuracy of the FSTDs d s racy of the skull match, the geometric deviation between measurement. When the surface contained noise and sharp the deformed template skull and JNS 1 was quantitatively features, e.g. boundary of the surface, normal estimation calculated and depicted in a graphical format. remained a challenge. The cost function was as follows: We extracted stable regions with robust normal estima- tion from the whole skull and then used FSTDs of the ver- () = () + () + () (2) l d s tices within these stable regions to accomplish FA. It com- prised two steps: firstly, we calculated the FSTDs of all the where α, β, and λ denoted the weights that guided the opti- vertices along the closest distance vectors (Huempfner-Hierl mization process. E (X) was used to initialize and guide the et al. 2015). For every vertex of the template skull, the near- registration as follows: est point on the template face was searched, and the FSTDs () = ‖ v − m ‖ were defined as the Euclidean distances between every (3) l i i i pair of corresponding vertices. Secondly, the discrepancies where m was the i-th landmark and semi-landmark of JNS 1 between FSTDs along the normal vectors and those along and v was the i-th corresponding landmark and semi-land- the closest distance vectors were calculated. Once the devia- mark of resulting deformation of the template skull via TPS. tion was less than the threshold, the vertex was suggested to The local affine regularization term expressed the dis- be a stable vertex. Figure 4a shows the stable regions (red tance between a vertex of JNS 1 and the corresponding points) and unstable regions (blue points). In addition, the vertex of the resulting deformation of the template skull as boundary vertices of the skull often were not considered follows: to belong to the stable region. They can be extracted from the triangle meshes based on the assumption that the one- () = dist rq , rp (4) d i i i i ring adjacent points of every boundary vertex cannot form a closed loop (Shui et al. 2020). Figure 4b shows the bound- where dist() denoted the distances between the correspond- ary vertices of the template skull (green points). Neither ing points of JNS 1 and the resulting deformation of the the unstable regions nor the boundary vertices were used to template skull, and denoted the reliability of the corre- generate the coarsely approximated face. spondences between these two meshes. We assumed that the nearest points between two meshes were the correspond- ences, denoted by rq and rp . In this procedure, the angles i i between normal vectors of the corresponding points and the Euclidean distance of the corresponding points can be used to improve reliability and reject the outliers. The stiffness term was applied to regularize the deforma- tion as follows: � � = ‖ − ∙ ‖ ( ) (5) s i j where ‖ ∙ ‖ was the Frobenius norm. and were the F i j transformations of neighboring vertices, which were con- nected by an edge that belonged to the resulting deformation of the template skull. G = diag(1,1,1,γ) denoted a weighting matrix. Dense FSTDs of the template During the acquisition of FSTDs of the template, the normal vector that was almost perpendicular to the surface of the bony structure was con- Fig. 4 Segmentation of template skull. (a) Stable vertices (red points) sidered to be the measurement direction. A ray that started at and unstable vertices (blue points). (b) Boundary vertices (green points) 1 3 Archaeol Anthropol Sci (2021) 13:186 Page 9 of 20 186 A coarse‑to‑fine facial approximation The overall shape of between all corresponding vertices and their centroid was the facial approximation can be coarsely produced based calculated. Then, PCA was conducted on the Procrustes on the dense FSTDs of the template using the following aligned coordinates to construct a facial shape tangent space. equation: In this shape space, every sample was represented by the average face and the linear combinations of PC scores and f = s + ∙ d i i i i (6) corresponding independent orthogonal PCs. Next, Student’s t test was carried out to verify the significant level of PC where f and s denoted the geometric coordinates of the i-th i i of interest between the modern faces and the approximated vertex of the approximated face and JNS 1, respectively. V faces. Finally, a visualization technique was used to inves- represented the normal vector of the i-th vertex of JNS 1, tigate the extent to which PC greatly explained the main and d denoted the soft tissue thickness of the corresponding patterns of morphological variation. In this process, two new vertex of the template. faces along the positive and negative PC of interest were The predictions of FSTDs and corresponding measure- generated as follows: ment directions were always inconsistent with the actual ones; thus, the approximated face would be unsmooth. In () = + 3 (8) i i addition, there always existed some voids, such as the eyes, nose, and cheeks. We employed a TPS function to warp the where and denoted the average face and weighting candidate face to the coarsely approximated face to improve coefficient (it was set to 1 or − 1), and denoted the stand- the approximation. Because the position of the control point ard deviation of the i-th PC, and represented the i-th PC. located on the two faces will greatly influence the deforma- tion, we calculated the corresponding intersection points as The effects of FSTDs and skull morphology control points based on the known anatomical landmarks on the approximated face and semi-landmarks of JNS 1. Additionally, we offered a tool to mathematically calcu- It is noted that skull morphology and the distribution of late multiple approximations that simulated the mandible FSTDs are the two fundamental components of FA. We morphology changes using the following equation: examined how the choice of FSTDs affected the approxi - mated face. The approximated faces of the same JNS 1 () = ∙ + (1 − ) ∙ (7) 1 2 were produced based on the average FSTDs of the females and males within our dataset. Then, the FSTDs deviation where represented the interpolated approxima- tion. and denoted the approximated faces of between two different approximations was calculated and 1 2 depicted in a graphical format. JNS 1 using Tabun 2 and JNS 1 using Mauer 1, respectively. ∈ [0,1] represented the weighting coefficient. As the real mandible of JNS 1 was not survived, we inves- tigated the effect of different mandibles on the approximated Evaluation of the reliability We validated the reliability of faces. Different approximated faces of JNS 1 using Tabun 2 and JNS 1 using Mauer 1 were produced based on the same the approximated face by means of examining whether or not the distribution of FSTDs of the template was consist- distributions of FSTDs, respectively. Then, the geometric deviation between the approximations was used to examine ent with that of the approximated face. The FSTDs defined along the closest distance vectors were used, because they the shape difference. were insensitive to measurement direction and data noise (Gietzen et al. 2019). The FSTDs deviation between the tem- Results plate and the approximation was calculated and visualized. Facial approximation of JNS 1 Morphological shape variations of facial approximation Since JNS 1 is suggested to be a female, the average skull GM was carried out to capture the main features of the approximated faces and examine the geometric morphologi- and face of the female group (Fig. 5a) was chosen as the template to approximate the facial appearance. Based on cal variations between the approximated faces and modern human faces. GPA was first used to register all the vertices 495 anatomical landmarks and semi-landmarks, we first used the TPS deformation approach to approximate the of the approximated faces and modern human faces, remov- ing translation, rotation, and scaling (O’Higgins and Jones facial appearance of JNS 1 using Tabun 2. Figure 5b and c show the deformed template skull and the candidate face. 1998). Thus, all the faces can be represented in the non-lin- ear Kendall’s shape space. The centroid size (CS) which is Figure 5d shows the template skull (gray color) and JNS 1 (peach color). It can be seen that the deformed template skull defined as the square root of the summed squared distances 1 3 186 Page 10 of 20 Archaeol Anthropol Sci (2021) 13:186 Fig. 5 The deformation of the template skull and face using TPS. deformed template skull (black points) and JNS 1(peach color). (f) (a) The template skull and face. (b) The deformed temple skull. (c) The visualization of geometric deviation between the deformed tem- The deformed template face. (d) The template skull (gray color) and plate skull and JNS 1 using Tabun 2 JNS 1 using Tabun 2 (peach color). (e) The superimposition of the (black points) does not match JNS 1 (peach color) closely, the corresponding vertices of JNS 1. The non-rigid registra- as shown in Fig. 5e. The geometric difference between tion was used to warp the deformed template skull to JNS 1 the deformed template skull and JNS 1 is calculated and to generate a better deformation result (Fig. 6a). Figure 6b depicted in Fig. 5f, where the Euclidean distance of every shows JNS 1 (peach color) and the deformation result (black corresponding vertex was calculated. The average distance points). The geometric difference between the deformed value of the vertex was 0.98 mm. The smaller deviation template skull and JNS 1 is calculated and visualized in regions were located around the top of the cranial vault, Fig. 6c, where the Euclidean distance of every correspond- whereas the largest geometric deviation regions were found ing vertex was calculated. Almost 96.9% of overall vertices around superciliary arches, the bottom of frontal, zygomatic on JNS 1 showed the deviation within a deformation error arch, lateral mastoid process, the bottom of the occipital of 1.0 mm, and the average deviation value was 0.54 mm. bone, partial regions of maxilla and teeth, and the greater The area (black points) where the deviation was greater than wing of the sphenoid. It is noted that the approximated face 2.0 mm can be mainly observed around the boundaries of the might be inaccurate around these regions, and the conven- bony structure. The region where the geometric deviation tional deformation-based FA approach needs to be improved. was greater can be observed around the superciliary arches We proposed a coarse-to-fine facial approximation and nasal bone. approach by attaching the dense FSTDs of the template to Fig. 6 The deformation of the template skull using the proposed (peach color). (c) The visualization of geometric deviation between method. (a) The deformed template skull. (b) The superimposition of the deformed template skull and JNS 1 using Tabun 2 the deformed template skull (black points) and JNS 1 using Tabun 2 1 3 Archaeol Anthropol Sci (2021) 13:186 Page 11 of 20 186 Dense FSTDs of the template that were calculated along ramus, greater wing of the sphenoid, and zygomatic. The the normal vector are visualized in Fig. 7a. It appeared that deviation of the vertex within the superciliary arches and the FSTDs were almost distributed symmetrically along the nasal bone was greater as well. Figure 8e shows the FSTDs midsagittal plane. The thinner FSTDs were observed at ver- deviation between the coarsely approximated face and the tices located around the forehead, superciliary arches, nasal template. Almost 87.7% of overall vertices on the coarse bone, and the cranial vault, and the thickest FSTDs were approximation showed the deviation within a discrepancy observed around the cheek region, the greater wing of the of 1.0 mm, and the average deviation value was 0.39 mm. sphenoid, and the bottom of the occipital bone. Figure 7b It can be seen that the distribution of average FSTDs of the and c display the point clouds and the triangle meshes of the template was very close to that of the coarse approximation, coarsely approximated face. For every landmark and semi- except for the side of the teeth and mandibular condyle. Fig- landmark of JNS 1, we calculate the corresponding points ure 8f illustrates the FSTDs deviation between the resulting on the candidate face (left figure) and the coarsely approxi- approximation and the template. Almost 76.4% of overall mated face (right figure), as seen in Fig. 7d. Figure 7e and f vertices on the improved approximation showed the devia- show the improved resulting approximation that is created tion within a discrepancy of 1.0 mm, and the average devia- by warping the candidate face to the coarse approximation tion value was 0.85 mm. The area where the deviation was using TPS. The result shows that the approximated face has greater than 2.5 mm can be observed around the coronoid a lower forehead, and a protruding, wider, and elongated process, both sides of the teeth and the greater wing of the middle and upper face. Also, it has robust eyebrows, a broad sphenoid. and short nose, and a wide mouth. Because the FSTDs are the fundamental basis of the We compared the approximated faces using two differ - facial approximation in our method, we employed the FSTDs ent methods based on the comparison of the distribution of that were calculated using different methods to predict the FSTDs. Figure 8a–c illustrate the distributions of FSTDs of overall shape of the coarsely approximated faces. Figure 9 the candidate face, the coarsely approximated face, and the shows the point clouds of the coarse approximation based improved resulting approximation, respectively. Figure 8d on the normal vectors of the overall vertices, rather than shows the FSTDs deviation between the candidate face and the stable vertices. It appeared that much noise and outliers the template (Table 3). Almost 62.9% of overall vertices occurred around the coarsely approximated face. Figure 10a on the candidate face showed geometric deviation within displays the FSTDs that are computed using a cylindrical a discrepancy of 1.0 mm, and the average deviation value sampling method (Shui et al. 2016). Figure 10b shows the was 1.17 mm. The area where the deviation was greater coarsely approximated faces that are recreated by attach- than 2.5 mm can be observed around the coronoid process, ing these FSTDs to JNS 1 along the cylindrical sampling Fig. 7 The facial approximation of JNS 1 using the proposed method. coarsely approximated face and the deformed face using TPS. (e) The (a) The visualization of dense FSTDs of the template. (b) The point improved resulting approximation by warping the deformed template clouds of the coarsely approximated face. (c) The surface of the face. (f) The superimposition of JNS 1 and the facial approximation coarsely approximated face. (d) The corresponding points on the 1 3 186 Page 12 of 20 Archaeol Anthropol Sci (2021) 13:186 Fig. 8 Comparisons of the facial approximations generated by differ - deviation between the template and the candidate face. (e) The ent methods. (a) The distribution of FSTDs of the candidate face. (b) FSTDs deviation between the template and the coarse approxima- The distribution of FSTDs of the coarse approximation. (c)The dis- tion. (f) The FSTDs deviation between the template and the improved tribution of FSTDs of the improved approximation. (d) The FSTDs approximation Table 3 FSTDs deviation FSTDs deviation The candidate face The coarsely approximated The improved approxi- (mm) and the percentage face mated face distribution (%) between three approximations and the % Deviation % Deviation % Deviation template (0.0, 0.5] 40.5% 0.23 80.6% 0.08 51.6% 0.22 (0.5, 1.0] 22.4% 0.73 7.1% 0.71 24.8% 0.71 (1.0, 2.5] 24.5% 1.59 8.5% 1.62 17.3% 1.51 > 2.5 12.6% 4.16 3.8% 3.74 6.3% 4.74 Total 100% 1.17 100% 0.39 100% 0.85 Fig. 9 The coarsely approxi- mated face using the normal vectors of overall the vertices vector. Figure 10c shows the FSTDs deviation between the a discrepancy of 1.0 mm, and the average deviation value template and the approximation. Almost 60.5% of overall was 1.3 mm. The area where the deviation was greater than vertices on the approximation showed the deviation within 2.5 mm can be observed around the side of teeth, greater 1 3 Archaeol Anthropol Sci (2021) 13:186 Page 13 of 20 186 Fig. 10 The coarsely approximated face using the FSTDs along the coarsely approximated face. (c) The FSTDs deviation between the cylindrical sampling vectors. (a) The distribution of average FSTDs template and the coarse approximation of the template using the cylindrical sampling method. (b) The wing of the sphenoid, the bottom of zygomatic region, distribution of the average FSTDs of the males is consistent mandibular condyle, and superciliary arches. These results with that of the 04 skull. However, the approximated chin indicated the proposed FA method can accurately assign the of the 04 skull bore little visual resemblance to the actual FSTDs to the corresponding vertices of JNS 1. chin, because the mental protuberance and tubercle seemed to be incomplete. Facial approximation of modern humans These results indicate that the overall shapes of the approximated faces bore resemblances to the actual faces. In order to evaluate the reliability of the proposed FA The area where the deviation was thinner can be observed approach, we approximated faces of two modern female around the forehead, nasal bone, and maxillary bone. But the skulls (01 and 03 skulls) and two modern male skulls (02 and greater deviation can be found around the cheeks, nose tip, 04 skulls). Then, we compare the FSTDs deviation between zygomatic region, parietal and temporal regions, and men- the approximated and actual faces, and the geometric varia- tal protuberance. The distribution of FSTDs deviation was tions between the approximated and actual faces, as shown relatively consistent with that of the geometric discrepancy, in Fig. 11. From left to right, the leftmost column displayed but the average value of the FSTDs deviation was less than the skull, and the next two columns illustrated the approxi- the geometric deviation. Thus, the proposed FA approach mated face and the distribution of FSTDs. The middle two can be applied to modern humans, and the FSTDs deviation columns displayed the actual face and the distribution of is a promising evaluation indicator to access the reliability FSTDs. The sixth column depicted the FSTDs deviation of the approximated face. for every vertex between the approximated and actual faces (Table 4). Additionally, the geometric deviation between the approximated and actual faces was calculated (Table 5). The effects of FSTDs and skull morphology on facial The rightmost column displayed the geometric discrepancy approximation of every vertex between the approximated and actual faces. Figure 11a shows the approximation of the 01 skull based on We can recreate multiple approximated faces based on the the average FSTDs of the females. Almost 59.3% of overall distributions of FSTDs of different templates. Figure 12a vertices showed the FSTDs deviation within 2.5 mm, and the shows another approximation of JNS 1 based on the average average deviation value was 2.12 mm. Figure 11b shows the FSTDs of the males within our dataset. Figure 12b displays approximation of the 02 skull based on the average FSTDs the FSTDs of the approximated face. The FSTDs deviation of the females. Almost 63.5% of overall vertices showed between two approximations based on the average FSTDs the FSTDs deviation within 2.5 mm, and the average devia- of the females and males is calculated, as shown in Fig. 12c. tion value was 2.19 mm. Figure 11c displays the approxi- Almost 69.9% of overall vertices showed the FSTDs devia- mation of the 03 skull based on the average FSTDs of the tion within a discrepancy of 1.0 mm, and the average FSTDs males. Almost 80.8% of overall vertices showed the FSTDs deviation value was 0.81 mm. The area (blue, red, and black deviation within a discrepancy of 2.5 mm, and the average points) where the FSTDs deviation was greater than 1.0 mm deviation value was 1.43 mm. Figure 11d shows the approxi- can be observed around the mouth, nasal base, zygomatic, mation of the 04 skull based on the average FSTDs of the parietal and occipital regions, and cheeks. males. Almost 94.4% of overall vertices showed the FSTDs Using the above FA method, we recreated the approxi- deviation within a discrepancy of 2.5 mm, and the average mated faces of JNS 1 using Mauer 1. Figure 13a shows deviation value was 0.95 mm. A potential reason is that the the approximated face based on the average FSTDs of 1 3 186 Page 14 of 20 Archaeol Anthropol Sci (2021) 13:186 Fig. 11 Four approximation examples of modern humans. From left of FSTDs. The rightmost two columns depicted the FSTDs devia- to right, the left column displayed the skull and the next two columns tion and geometric discrepancy between the approximated and actual illustrated the approximated face and the distribution of FSTDs. The faces. (a) The 01 skull. (b) The 02 skull. (c) The 03 skull. (d) The 04 middle two columns displayed the actual face and the distribution skull Table 4 FSTDs deviation (mm) FSTDs deviation 01 skull 02 skull 03 skull 04 skull and the percentage distribution (%) between the approximated % Deviation % Deviation % Deviation % Deviation and actual faces (0.0, 1.0] 29.8% 0.44 17.0% 0.52 52.1% 0.42 63.8% 0.47 (1.0, 2.5] 30.5% 1.73 46.5% 1.84 28.7% 1.64 30.6% 1.57 (2.5, 5.0] 36.4% 3.40 33.7% 3.17 16.9% 3.68 5.6% 3.02 > 5.0 3.3% 5.73 2.8% 6.40 2.3% 5.35 - - Total 100% 2.09 100% 2.19 100% 1.43 100% 0.95 Table 5 Geometric deviation Geometric deviation 01 skull 02 skull 03 skull 04 skull (mm) and the percentage distribution (%) between the % Deviation % Deviation % Deviation % Deviation approximated and actual faces (0.0, 1.0] 27.4% 0.57 16.9% 0.57 52.3% 0.47 53.0% 0.59 (1.0, 2.5] 33.9% 1.71 44.9% 1.81 29.4% 1.61 35.8% 1.57 (2.5, 5.0] 33.4% 3.41 32.5% 3.28 15.6% 3.75 9.1% 3.16 > 5.0 5.3% 6.44 5.7% 6.78 2.7% 6.46 2.1% 6.58 Total 100% 2.21 100% 2.36 100% 1.49 100% 1.30 1 3 Archaeol Anthropol Sci (2021) 13:186 Page 15 of 20 186 Fig. 12 Facial approximation based on the average FSTDs of the males. (a) The approximated face. (b) The distribution of FSTDs of the approx- imation. (c) The FSTDs deviation between two approximations the females. Figure 13b shows the geometric variations Shape analysis of the approximated faces between this approximation and the approximation of JNS 1 using Tabun 2. We also use the distribution of average GM analysis was conducted to capture the geometric FSTDs of the males to recreate an approximation of JNS 1 features of the approximated faces. A total of 27 PCs using Mauer 1, as shown in Fig. 13c. Figure 13d shows the accounted for over 95% of the morphological variance in geometric deviation between this approximation and the the shape space. The first PC (PC 1) accounted for 31.6% approximation of JNS 1 using Tabun 2. The approximation of the morphological variance, and the second PC (PC of JNS 1 using Mauer 1 has a wider face and a more robust 2) accounted for 13.3% of the morphological variance. chin than the approximations of JNS 1 using Tabun 2. Figure 15 shows the plots of the first two PCs of four dif- These results showed that the geometric shape of the skull ferent approximations (Fig. S1), including the approxi- greatly influences the overall shape of the approximated mated face that used the average FSTDs of the females and faces. Based on the distribution of average FSTDs of the Tabun 2 (green point), the approximated face that used the females, we employed the linear interpolation method to average FSTDs of the males and Tabun 2 (yellow point), mathematically approximate three faces. Figure 14 dis- the approximated face that used the average FSTDs of the plays three approximations using η = 0.25, η = 0.50, and females and Mauer 1 (black point), the approximated face η = 0.75. that used the average FSTDs of the males and Mauer 1 (cyan point), 30 modern female faces (red points), and 30 Fig. 13 Facial approximation of JNS 1 using Mauer 1. (a) The approximated face using aver- age FSTDs of the females. (b) Geometric deviation between two approximated faces. (c) The approximated face using the average FSTDs of the males. (d) Geometric deviation between two approximated faces 1 3 186 Page 16 of 20 Archaeol Anthropol Sci (2021) 13:186 Fig. 14 Three approximated faces using linear interpolation. (a) η = 0.25. (b) η = 0.50, (c) η = 0.75 Fig. 15 Comparison of the approximated faces and modern human approximated face (black point) that used Mauer 1 and the FSTDs faces in the shape space. Scatterplots of 30 female faces (red points), of the females, the approximated face (cyan point) that used Mauer 1 30 male faces (blue points), the approximated face (green point) that and FSTDs of the males. The frontal and profile views of new gener - used Tabun 2 and the FSTDs of the females, the approximated face ated faces corresponding to the extreme limits of PC 1 and PC 2 (yellow point) that used Tabun 2 and the FSTDs of the males, the modern male faces (blue points). It is of note that PC 1 narrower mouth, nose, and chin. All the approximated (p < 0.005) and PC 2 (p < 0.05) have significant differences faces were located at the extreme positive end of PC 1. It between modern human faces and the approximated faces. indicated that the approximated face was greatly different To identify the main patterns of shape variance, four new from modern human faces and verified the characteristic faces were recreated along the positive and negative direc- features of the approximations. When the same FSTDs tions of PC 1 and PC 2. The positive PC 1 connected with were performed, the approximated faces that used Mauer the approximated face with a relatively lower forehead, 1 were located on the right side along the positive PC 1. and robust and wide eyebrows; a protruding, wider, and It indicates that these approximations have wider cheeks elongated middle and upper face; a broad and short nose; and robust chins. a wider mouth; and a robust chin. By contrast, the negative PC 1 represented a face with a prominent and protruding forehead, a narrower middle and upper face, and relatively 1 3 Archaeol Anthropol Sci (2021) 13:186 Page 17 of 20 186 confidently reasonable for JNS 1. This study considers the Discussion distribution of average FSTDs of modern humans as that of the archaic human and strongly suggests that FSTDs of Many previous studies have employed manual and comput- the reliable regions along the normal vectors are appro- erized FA approaches to modern humans in archaeology priate for approximating the facial appearances of archaic and anthropology (Claes et al. 2010; Wilkinson 2010). humans. However, when FA is applied to archaic humans, the main It is worth mentioning that the FSTDs make a great con- challenges are the poor preservation of archaic human fos- tribution to improving the accuracy and reliability of the sils and the absence of their craniofacial relationship and approximated face (Claes et al. 2006; Starbuck and Ward anatomical knowledge. In this study, we propose a comput- 2007). It is acknowledged that many factors, such as sex, erized coarse-to-fine FA approach to attach the distribution age, nutrition status, body mass index, and ethnic groups, of average FSTDs of modern humans to archaic humans. will impact the FSTDs of landmarks. Previous studies The resulting approximation is promising, objective, and always placed the landmarks on the skull manually and repeatable, and the reliability can be evaluated through then acquired the FSTDs at these landmarks (Stephan 2017). a quantitative comparison of the distributions of FSTDs. We investigated the effects of the distribution of FSTDs on Furthermore, we investigate the effects of skull geometry FA. Approximated faces of the same JNS 1 using different and the distribution of FSTDs on the approximated face. FSTDs shared a resemblance. The larger deviation areas of The first stage of FA is to examine and restore the dry two approximated faces between using FSTDs of the males skull. When the skull had missing parts or distortion, the and females were almost consistent with the regions of the TPS function was always used to deform the reflection FSTDs discrepancy between males and females (Shui et al. structure of the intact side to fill in the gaps based on a 2016). Once a reliable and verified craniofacial relationship landmark and semi-landmark configuration. Other stud- can be obtained, a more reliable and accurate approximation ies attempted to use computerized approaches to virtually would be recreated. reassemble the skull fractures together (Yu et al. 2012). Because of the greater shape variations between the In the worst case, the fossil specimen, as in the case of the modern and archaic human skulls, it remains challenging to mandible of JNS 1, had not survived. In these cases, there assign the FSTDs of the template to those of JNS 1. Previous is a potential solution, using a mandible that has similar studies often employed the deformation-based FA approach age and morphological features to match the cranium. But (Nelson and Michael 1998; Turner et al. 2005; Deng et al. since the mandibles of archaic humans are rarely found, 2011). It is accepted that the closer the deformed template and the geometric shapes of the mandibles are always and the dry skulls matched, the more confidence there was unique, it remains challenging to select an appropriate in the reliability of the approximated face. However, the mandible to provide a good fit with the JNS 1 cranium. performance of the TPS deformation even incorporating a We attempted to use different mandibles to match the JNS regularization is inadequate accuracy when two skulls have 1 cranium and recreated different approximations based great geometric differences. To address this problem, this on the repaired skull. According to these approximated study employed a hybrid non-rigid registration algorithm to facial appearances, multiple approximations can further establish a high-quality set of geometric correspondences. be mathematically recreated to provide some references Additionally, this approach can make use of FSTDs for every to describe the overall shape. Although these approxi- sample within the dataset to recreate a range of multiple mations cannot be interpreted anatomically, they might approximated faces and then use PCA to construct a tailored provide a new perspective for researchers to illustrate the approximation-space for JNS 1. In this context, the missing facial appearance. areas of the coarsely approximated face can be better pre- During FA, the prediction of the craniofacial relation- dicted (Gietzen et al. 2019), and a range of possible approxi- ship between soft tissue and the dry skull and the assign- mation can be recreated by using appropriate coefficients of ment of the predicted craniofacial relationship are two PCs of interests (Shui and Wu 2018). fundamental questions. The average FSTDs at landmarks As previous studies mentioned (Oxnard and O’Higgins and muscle structures of modern humans are always sug- 2009; Wärmländer et al. 2019), the anatomical landmarks gested to be the craniofacial relationship of people in the and semi-landmarks should be very carefully designed with past (Hamre et al. 2017). In the study previously men- regard to the research question. The purpose of this study is tioned, the FSTDs of chimpanzees can be used to depict to approximate the overall shape of the facial appearance. a thinner mid-face of archaic humans because FSTDs Previous studies employed different numbers of landmarks around the cheek were almost half that of modern humans (Vandermeulen et al. 2006; Deng et al. 2011) and topographic (Hayes et al. 2013). However, due to the lack of evidence, features, e.g. crest lines (Turner et al. 2005), to guide the defor- there is a particular challenge to decide which FSTDs are mation. But there is no standard criterion for performing FA. 1 3 186 Page 18 of 20 Archaeol Anthropol Sci (2021) 13:186 We recommend that landmarks and semi-landmarks need to archaic human. This study proposed a coarse-to-fine FA approach cover the entire skull, particularly around the region where the based on dense FSTDs of modern humans and presented an template and dry skulls are quite different. Such definitions evaluation approach to validate the reliability of FA. We also will improve the certainties of the establishment of dense geo- investigated the effects of skull morphology and the distribution metric correspondences and so enhance the reliability of the of FSTDs on the approximated faces. Since the mandibles of approximated face. Additionally, we employed GM to capture archaic humans are rarely found, it remains challenging to select the main features of the approximated faces. The dense cor- an appropriate mandible for the JNS 1 cranium. In the future, we responding vertices are used to provide better visualization will attempt to collect the different mandibles of archaic humans and interpretation, rather than the use of landmarks and semi- to improve the reliability of the approximation of JNS 1. landmarks. But due to the complexity of biological structures, Supplementary Information The online version contains supplemen- we need to carefully examine the effect of landmarks, semi- tary material available at https://doi. or g/10. 1007/ s12520- 021- 01450-w . landmarks, and high-density correspondences on the shape analysis and then decide which type of corresponding points Acknowledgements The authors would like to thank all anonymous can provide a reliable interpretation. reviewers. We are grateful to Prof. Julian Richards for giving valu- The geometric comparison of the actual and approxi- able comments and feedback. We acknowledge and thank Dr. Antonio Profico for providing access to the Mauer 1 mandible. This work is mated faces has been used to validate the reliability of FA. supported by the Strategic Priority Research Program (XDB26000000). In this process, the shell-to-shell deviation, surface-to-sur- face deviation, and the shortest distance between two point Open Access This article is licensed under a Creative Commons Attri- clouds have been conducted (Wilkinson et al. 2006; Short bution 4.0 International License, which permits use, sharing, adapta- tion, distribution and reproduction in any medium or format, as long et al. 2014; Miranda et al. 2018). But, the resulting reg- as you give appropriate credit to the original author(s) and the source, istration impacts the comparisons of discrepancy between provide a link to the Creative Commons licence, and indicate if changes the approximated and actual faces. Additionally, since no were made. The images or other third party material in this article are verified actual faces can be provided, the comparison of the included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in actual and approximated faces cannot be used to archaic the article's Creative Commons licence and your intended use is not humans. Thus, the reliability of the approximated face is permitted by statutory regulation or exceeds the permitted use, you will mainly evaluated based upon the experience and knowledge need to obtain permission directly from the copyright holder. To view a of the experts. Even though the experts marked the areas of copy of this licence, visit http://cr eativ ecommons. or g/licen ses/ b y/4.0/ . the approximated faces that need to be improved, research- ers still do not know how to revise them exactly. To tackle this problem, we can integrate the approximation of facial appearance and recognition of the less confidence in approx- References imated regions iteratively. This method provides a promising tool to allow researchers to examine the extent to which the Andrews P (1986) Fossil evidence on human origins and dispersal. Cold Spring Harb Symp Quant Biol 51:419–428 resulting approximation is unreliable and where it needs to Athreya S, Wu X (2017) A multivariate assessment of the Dali hominin cra- be further revised to improve the reliability. nium from China: morphological affinities and implications for Pleisto- However, uncertainty still remains since there are no cene evolution in East Asia. Am J Phys Anthropol 164:679–701 actual archaic human faces and we cannot know the actual Benazzi S, Fantini M, Crescenzio FD, Mallegni G, Mallegni F, Per- siani F, Gruppioni G (2009) The face of the poet Dante Alighieri relationship between soft tissue and skull of archaic human. reconstructed by virtual modelling and forensic anthropology Also, it remains challenging to understand the relationship techniques. J Archaeol Sci 36:278–283 between facial features, e.g. eyes, nose, mouth, and ears, Berar M, Tilotta FM, Glaunès JA, Rozenholc Y (2011) Craniofacial and bony structures. Thus, the approximated features need reconstruction as a prediction problem using a latent root regres- sion model. Forensic Sci Int 210:228–236 to be further improved, e.g. the size and orientation of nasal Bookstein FL (1989) Principal warps: thin-plate splines and the decom- aperture. In addition, the ages and features of Tabun 2 and position of deformations. IEEE T Pattern Anal 11:567–585 Mauer 1 are different from JNS 1, and an appropriate mandi- Claes P, Vandermeulen D, De Greef S, Willems G, Suetens P (2006) ble needs to be used to enhance the reliability of the approxi- Craniofacial reconstruction using a combined statistical model of face shape and soft tissue depths: methodology and validation. mated face. Forensic Sci Int 159:S147–S158 Claes P, Vandermeulen D, De Greef S, Willems G, Clement JG, Suetens P (2010) Computerized craniofacial reconstruction: con- Conclusion ceptual framework and review. Forensic Sci Int 201:138–145 De Buhan M, Nardoni C (2018) A facial reconstruction method based on new mesh deformation techniques. Forensic Sci Res 3(3):256–273 The approximated face of JNS 1 is a typical case of interdiscipli- Deng Q, Zhou M, Shui W, Wu Z, Ji Y, Bai R (2011) A novel skull reg- nary study that provides anthropologists and the general public istration based on global and local deformations for craniofacial with an improved visual interpretation of the facial morphology of reconstruction. Forensic Sci Int 208:95–102 1 3 Archaeol Anthropol Sci (2021) 13:186 Page 19 of 20 186 Deng Q, Zhou M, Wu Z, Shui W, Ji Y, Wang X, Liu CYJ, Huang Y, Oxnard CE, O’Higgins P (2009) Biology clearly needs morphometrics. Jiang H (2016) A regional method for craniofacial reconstruc- Does morphometrics need biology? Biol Theory 4:84–97 tion based on coordinate adjustments and a new fusion strategy. Paysan P, Knothe R, Amberg B, Romdhani S, Vetter T (2009) Face Forensic Sci Int 259:19–31 reconstruction from skull shapes and physical attributes. Lecture Fatuzzo G, Sequenzia G, Oliveri SM (2016) Virtual anthropology Notes in Computer Science, Pattern Recognition. Springer, Berlin, and rapid prototyping: a study of Vincenzo Bellini’s death masks pp 232–241 in relation to autopsy documentation. Digital Applications in Quam RM, Smith FH (1998) A reassessment of the Tabun C2 mandible. Archaeology and Cultural Heritage 3:117–125 In: Akazawa T, Aoki K, Bar-Yosef O (eds) Neandertals and Modern Gietzen T, Brylka R, Achenbach J, Hebel ZK, Schömer E, Botsch Humans in Western Asia. Springer, New York, pp 405–421 M, Schwanecke U, Schulze R (2019) A method for automatic Quatrehomme G, Cotin S, Subsol G, Delingette H, Garidel Y, Grévin forensic facial reconstruction based on dense statistics of soft G, Fidrich M, Bailet P, Ollier A (1997) A fully three-dimensional tissue thickness. PLoS One 14(1):e0210257 method for facial reconstruction based on deformable models. J Gunz P, Mitteroecker P (2013) Semilandmarks: a method for quan- Forensic Sci 42:649–652 tifying curves and surfaces. Hystrix 24(1):103–109 Rightmire GP (1998) Human evolution in the Middle Pleistocene: The Gunz P, Mitteroecker P, Neubauer S, Weber GW, Bookstein FL role of Homo heidelbergensis. Evol Anthropol 6:218–227 (2009) Principles for the virtual reconstruction of hominin cra- Rosenberg KR, Zuné L, Ruff CB (2006) Body size, body proportions, nia. J Hum Evol 57:48–62 and encephalization in a Middle Pleistocene archaic human from Hamre SS, Ersland GA, Daux V, Parson W, Wilkinson C (2017) northern China. PNAS 103:3552–3556 Three individuals, three stories, three burials from medieval Schwartz J, Tattersall I (2000) The human chin revisited: what is it and Trondheim. Norway. PLoS One 12(7):e0180277 who has it? J Hum Evol 38(3):367–409 Harvati K, Nicholson Lopez E (2017) A 3-D look at the Tabun C2 jaw. In: Short LJ, Khambay B, Ayoub A, Erolin C, Rynn C, Wilkinson C (2014) Marom A, Hovers E (eds) Human paleontology and Prehistory: 203– Validation of a computer modelled forensic facial reconstruction 214, vertebrate paleobiology and paleoanthropology. Springer, Berlin technique using CT data from live subjects: a pilot study. Forensic Hayes S (2016) Faces in the museum: revising the methods of facial Sci Int 237:147.e1-147.e8 reconstructions. Museum Management and Curatorship 31:218–245 Shui W, Wu X (2018) Three-dimensional craniofacial reconstruction Hayes S, Buckley H, Bradley R, Milne N, Dennison J (2012) Approximating the of an ancient Qihe human skull. Chin Sci Bull 63:745–754 (in face of ‘aunty’: a question of likeness. J Archaeol Method Th 19:306–321 Chinese) Hayes S, Sutikna T, Morwood M (2013) Faces of Homo floresiensis Shui W, Zhou M, Deng Q, Wu Z, Ji Y, Li K, He T, Jiang H (2016) Densely (LB1). J Archaeol Sci 40:4400–4410 calculated facial soft tissue thickness for craniofacial reconstruction Hayes S, Shoocongdej R, Pureepatpong N, Sangvichien S, Chintakanon in Chinese adults. Forensic Sci Int 266:573.e1-573.e12 K (2017) A Late Pleistocene woman from Tham Lod, Thailand: Shui W, Zhou M, Maddock S, He T, Wang X, Deng Q (2017) A PCA- the influence of today on a face from the past: the influence of based method for determining craniofacial relationship and sexual today on a face from the past. Antiquity 91(356):289–303 dimorphism of facial shapes. Comput Biol Med 90:33–49 Herrera RJ, Garcia-Bertrand R (2018) Ancestral DNA, human origins, Shui W, Zhou M, Maddock S, Ji Y, Deng Q, Li K, Fan Y, Li Y, Wu X and migrations. Academic Press (2020) A computerized craniofacial reconstruction method for an Hublin JJ (2013) The Middle Pleistocene record: on the ancestry of unidentified skull based on statistical shape models. Multimedia Neandertals, modern humans and others. A companion to paleo- Tools and Application 79:25589–25611 anthropology. Blackwell Publishing Ltd, Oxford, pp 517–537 Starbuck JM, Ward RE (2007) The affect of tissue depth variation on Huempfner-Hierl H, Bohne A, Schaller A, Wollny G, Hierl T (2015) craniofacial reconstructions. Forensic Sci Int 172:130–136 Does facial soft tissue protect against zygomatic fractures? Results Stephan CN (2017) 2018 tallied facial soft tissue thicknesses for adults of a finite element analysis. Head Face Med. 11:21 and sub-adults. Forensic Sci Int 280:113–123 Kustar A (2004) The facial restoration of Antal Simon, a Hungar- Turner WD, Brown RE, Kelliher TP, Tu PH, Taister MA, Miller KW ian priest-teacher of the 19th c. HOMO-Journal of Comparative (2005) A novel method of automated skull registration for forensic Human Biology 55:77–90 facial approximation. Forensic Sci Int 154:149–158 Lee WJ, Yoon AY, Mi KS, Wilkinson CM, Dong HS (2014) The archaeo- Vanezis P, Blowes RW, Linney AD, Tan AC, Richards R, Neave R (1989) logical contribution of forensic craniofacial reconstruction to a portrait Application of 3-D computer graphics for facial reconstruction and drawing of a Korean historical figure. J Archaeol Sci 49:228–236 comparison with sculpting techniques. Forensic Sci Int 42(1–2):69–84 Lorensen WE, Cline HE (1987) Marching cubes: a high-resolution 3D surface Verzé L (2009) History of facial reconstruction. Acta Biomed construction algorithm. SIGGRAPH Comput Graph 21(4):163–169 80(1):5–12 Marić J, Bašić Z, Jerković I, Mihanović F, Anđelinović Š, Kružić I Wagner GA, Krbetschek M, Degering D, Bahain JJ, Shao Q, Falguères (2020) Facial reconstruction of mummified remains of Christian C, Voinchet P, Dolo JM, Garcia T, Rightmire GP (2010) Radio- Saint-Nicolosa Bursa. J Cult Herit 42:249–254 metric dating of the type-site for Homo heidelbergensis at Mauer. Martin R (1928) Lehrbuch der Anthropologie in systematischer Germany. PNAS 107(46):19726–19730 Darstellung. G. Fischer, Stuttgart Wärmländer SK, Garvin H, Guyomarc’h P, Petaros A, Sholts SB Miranda GE, Wilkinson C, Roughley M, Beaini TL, Melani RFH (2018) (2019) Landmark typology in applied morphometrics studies: Assessment of accuracy and recognition of three-dimensional com- what’s the point?". Anat Rec 302(7):1144–1153 puterized forensic craniofacial reconstruction. PloS one 13:e0196770 Wiley DF, Amenta N, Alcantara DA, Ghosh D, Kil YJ, Delson E, Mounier A, Marchal F, Condemi S (2009) Is Homo heidelbergensis Harcourt-Smith W, Rohlf FJ, St John K, Hamann B (2005) Evo- a distinct species? New insight on the Mauer mandible. J Hum lutionary morphing. IEEE Visualization (VIS'05), pp 431–438. Evol 56:219–246 Wilkinson C (2005) Computerized forensic facial reconstruction: a Nelson L, Michael S (1998) The application of volume deformation to review of current systems. Forensic Sci Med Pathol 1:173–177 three-dimensional facial reconstruction: a comparison with previ- Wilkinson C (2010) Facial reconstruction—anatomical art or artistic ous techniques. Forensic Sci Int 94(3):167–181 anatomy? J Anat 216:235–250 O’Higgins P, Jones N (1998) Facial growth in Cercocebus torquatus: Wilkinson C, Rynn C, Peters H, Taister M, Kau CH, Richmond S (2006) A blind accuracy assessment of computer-modeled an application of three-dimensional geometric morphometric tech- niques to the study of morphological variation. J Anat 193:251–272 1 3 186 Page 20 of 20 Archaeol Anthropol Sci (2021) 13:186 forensic facial reconstruction using computed tomography data Publisher's note Springer Nature remains neutral with regard to from live subjects. Forensic Sci Med Pathol 2:179–187 jurisdictional claims in published maps and institutional affiliations. Wu R (1988) The Reconstruction of the fossil human skull from Jin- niushan, Yingkou, Liaoning Provience, and its main features. Acta Anthropologica Sinica 7:97–102 (in Chinese) Yu W, Li M, Li X (2012) Fragmented skull modeling using heat ker- nels. Graph Models 74(4):140–151 1 3
Archaeological and Anthropological Sciences – Springer Journals
Published: Nov 1, 2021
Keywords: Facial approximation; Archaic humans; Facial soft tissue thickness depths; Geometric correspondences; Assessment
You can share this free article with as many people as you like with the url below! We hope you enjoy this feature!
Read and print from thousands of top scholarly journals.
Already have an account? Log in
Bookmark this article. You can see your Bookmarks on your DeepDyve Library.
To save an article, log in first, or sign up for a DeepDyve account if you don’t already have one.
Copy and paste the desired citation format or use the link below to download a file formatted for EndNote
All DeepDyve websites use cookies to improve your online experience. They were placed on your computer when you launched this website. You can change your cookie settings through your browser.