May 2022 – October 2024
The issue of unidentified remains that cannot be traced back to their families or communities is an increasing humanitarian concern. Forensic facial reconstruction given a skull can be employed to support the identification of unidentified human remains. The aim of this research is to explore the possibility of using recent computer vision techniques, more specifically image-to-image translation models, to predict the 2D facial image of a human from their skull image. This by no means will be considered as a tool for identity confirmation but rather to generate leads to the possible identity. The document describes the problem definition in connection to existing computer vision models, the proposed solution, and the basic challenges.
EPFL PI: Prof. Dr. Amir Zamir
Partner: Denise Abboud (ICRC)
Photo: Copyright: ICRC./ Aleksander Imedashvili / 14.09.2016/ Tbilisi, Levan Samkharauli National Forensics Bureau. 17 sets of human remains of persons unaccounted for in relation to the 1992-1993 armed conflict in Abkhazia are handed over to their families… /V-P-GE-E-00825
Publications and/or scientific communication
- Publication on machine learning tool for generic robust 2D-3D translation: 4M-21: An Any-to-Any Vision Model for Tens of Tasks and Modalities, R. Bachmann*, O. Kar*, D. Mizrahi*, Ali Garjani, M. Gao, D. Griffiths, J. Hu, A. Dehghan, A. Zamir. In NeurIPS, 2024./ And webpage
- Publication on machine learning tool for generic robust 2D-3D translation: 3D Common Corruptions and Data Augmentation, O. Kar, T. Yeo, A. Atanov, A. Zamir. In CVPR, 2022. 3DCC paper and webpage
