The availability of image databases is of paramount importance in several Computer Vision and Pattern recognition problem, such as face identification, expression analysis and beauty machine analysis. While, 2D databases are easily accessible, only few 3D face (and body) datasets are available. In addition, to our knowledge, a comprehensive list of them has not been published in literature yet. Therefore, in this page, we review the existing databases of 3D faces (and bodies), along with available information regarding scanning device, resolution, precision and annotations.
Three scans, providing different orientations, of 120 subjects. The 3D models have been obtained with a proprietary system based on structured light. The output is a point cloud without texture. The quality is relatively low. No annotations are available.
The BFM is a PCA based morphable model, consisting of 53,490 vertices. New faces can be created as a linear combination of its 199 principal components. The coefficient for creating the 3D models of the faces contained in the FERET and CMU-PIE 2D databases are available. Two additional standard sets of landmarks (Farkas and MPEG4 FDP points) are also provided for each face.
Paysan, P., Knothe, R., Amberg, B., Romdhani, S., and Vetter, T. 2009. “A 3D Face Model for Pose and Illumination Invariant Face Recognition”. Proc. Sixth IEEE international Conference on Advanced Video and Signal Based Surveillance, AVSS 2009. pp. 296-301
Up to 54 scans in various poses, expressions and occlusion conditions of 105 different subjects (for a total of 4666 face scans). The DB includes 29 professional actors/actresses. 24 facial landmark points have been manually labeled. The sensor resolution is about 0.3 mm in each direction; high resolution color texture images (1600x1200 pixels) are provided.
Savran A., Alyüz N., Dibeklioğlu H., Çeliktutan O., Gökberk B., Sankur B., Akarun L., “Bosphorus Database for 3D Face Analysis”, The First COST 2101 Workshop on Biometrics and Identity Management (BIOID 2008), Roskilde University, Denmark, May 2008
The DB includes 25 scans of 100 subjects, each with a different facial expression, resulting in a total of 2500 3D models with an associated high resolution color texture (1040×1329 pixels). Each face scan has approximately 35,000 vertices. No annotations are available.
Yin L.; Chen X.; Sun Y.; Worm T.; Reale M., ”A High-Resolution 3D Dynamic Facial Expression Database”. Proc. 8th International Conference on Automatic Face and Gesture Recognition (FGR08), 17-19 September 2008
The database includes 500 faces, 250 male and 250 female Chinese subjects, with a natural expression and without glasses or other accessories, acquired with a CyberWare3030RGB/PS scanner. Model resolution is about 65,000 vertices. Texture is encoded as an RGB value per vertex. No annotations are available.
This is the only large DB available of body scans, and it contains 5000 full body textured scans. Anthropometric landmarks are present on the whole body. This is a commercial DB.
CASIA HFB includes up to 2 face images for 100 subjects, for a total of 192 scans. The models are textured and their resolution is 1 point every 3 mm in each direction. This makes about 3000 vertices for the face. No annotations are available.
Li S.Z., Lei Z., Ao M., “The HFB Face Database for Heterogeneous Face Biometrics Research”. In 6th IEEE Workshop on Object Tracking and Classification Beyond and in the Visible Spectrum (OTCBVS, in conjunction with CVPR 2009). Miami, Florida, June, 2009
The Twins Expression Challenge data set contains 3D face scans for 107 pairs of twins (214 individuals) each with 2 scans providing different expressions. The scans were acquired with a Minolta Vivid 910. No annotations are available
Two datasets are available. (1) Collection D. 953 different textured 3D scans from 277 human subjects, captured with a Minolta Vivid 900 3D range scanner. The output resolution is 640 x 480 pixels (3D data). No annotations are available. (2) Collection NDOff-2007. 6940 3D textured models of 387 human subjects taken with different head rotations (yaw and pitch angle are available).
Chang K., Bowyer K. W., and Flynn P. J., “Face recognition using 2D and 3D facial data,” ACM Workshop on Multimodal User Authentication, pp.25-32, December 2003
RGB-D images of 52 subjects (14 females, 38 males) caputured with the Kinect sensor at different time period and in different orientation, facial expression, lighting and occlusion conditions. Each image is annotated with the position of six landmarks.
Huynh T., Min R., Dugelay J., "An Efficient LBP-based Descriptor for Facial Depth Images applied to Gender Recognition using RGB-D Face Data," in ACCV Workshop on Computer Vision with Local Binary Pattern Variants 2012, Daejeon, Korea, November 2012
Several facial expressions of 466 subjects, for a total of 4007 3D scans, acquired with a Minolta Vivid 900 range scanner, with a resolution of 640x480 3D points. No annotations are available..
Phillips P., Flynn P., Scruggs T., Bowyer K., Chang J., Hoffman K., Marques J., Min J., and Worek W.. “Overview of the face recognition grand challenge”. In Proc. CVPR, pages 947–954, 2005
61 different subjects and 9 images with different orientation and expressions. The total number of 3D facial surfaces is 549. The 3D scans are not textured. Resolution is about 15,000 vertices. No annotations are available.
Sanchez A.. “GavabDB: A 3D Face Database”. Proc. 2nd COST Workshop on Biometrics on the Internet: Fundamentals, Advances and Applications, pp. 77-82, 2004
A total of more than 3000 3D scans have been acquired with different systems. 30 craniofacial anthropometric landmarks have been manually labeled on each model.
Evison MP, Vorder Bruegge RW. “The magna database: a database of three-dimensional facial images for research in human identification and recognition”. Forensic Sci Commun 2008;10(2)
200 laser-scanned (Cyberware TM) heads without hair. Textured models, resolution about 90,000 vertices. No annotations are available.
3187 3D images of 453 subjects capyured for each subject in different time periods. The 3D surfaces are obtained from a photometric stereo reconstruction which uses four differently lit colour photograph of size 1280x1024 pixels. Images are annotated with 11 facial landmarks.
1149 pairs of facial color and range images of 105 adult human subjects, acquired using a stereo imaging system manufactured by 3Q Technologies. Spatial resolution is about 0.32 mm along the x, y, and z axis. For each model, the manually located position of 25 landmarks is available.
Gupta S., Markey M. K., Bovik A. C., "Anthropometric 3D Face Recognition", International Journal of Computer Vision, 2010, Volume 90, 3:331-349
1473 scans of 143 subjects (98 male, 45 female), showing 4 expressions (neutral, smiling, bored, hungry) and acquired with a Minolta Vivid 900. 590 scans are occluded by various objects (eyeglasses, hands, hats, …). Each face is annotated with 7 landmarks, when visible, and, if occluded, with the the mask of its visible part
The database currently contains over 5000 3D face models of approximately 350 people (15 scans each, providing different postures and different orientations). No annotations are available. At the moment we are writing, the distribution of the DB has been suspended.
200 different subjects. Texture resolution is 296x347 pixel, 3D surfaces contain approximately 75,000 vertices. These scans have been used to develop the 3D morphable model (3DMM) of Vetter and Blanz, see reference below, which is distributed along with the DB. No annotations are available. However, the dense point to point correspondences between faces provided by the 3DMM, allows to easily obtain annotations on all samples by labeling a set of landmarks on a single face.
Blanz V. and Vetter T., “ A morphable model for the syntesis of 3D faces, “ SIGGRAPH 99 Proc, pp.187-194, 1999
UDHB is the general name of a collection of different DBs. They consist of several 3D textured scans of different subjects, taken in different poses, expressions and illumination conditions. Several scanning devices were also used. No annotations are available.
Four 3D head models of 295 subjects were acquired using a high precision stereo-based 3D camera; data are available in VRML format (about 4,500 vertices), with a texture image of 720x576 pixels. No annotations are available.
Messer K., Matas J, Kittler J, et al. Xm2vtsdb, “The extended m2vts database”. In International Conference of AVBPA, March 1999
Copyright © 2007 - All Rights Reserved
CG&VC @ Politecnico di Torino