3D face shapes are increasingly used for distraction detection in advanced driver-assistance systems (ADAS). However, the detection performance is hindered by the low signal quality due to the limitation of the in-vehicle video capturing technology. This project aims to train a neural-network-based generative model that improves the resolution of 3D face shape collected by in-vehicle camera for a diverse population. Massive 3D face shape data of adults will be utilized to pre-train a generative model for adults, which will be further extended to underrepresented population groups (e.g., children) via domain adaptation. A specific closed set semi-supervised domain adaptation algorithm will be developed to achieve a high reconstruction accuracy of 3D face shapes.
Funding: $60K (2023)
Goal: Build a generative AI model to reconstruct high resolution 3D face shapes from low resolution ones for children based on the sufficient training samples of adults.
Token Investors: Wenbo Sun, Arpan Kusari, Byoung-Keon (Daniel) Park, and David LeBlanc
Project ID: 1105