Zihua Liu

Zihua Liu

MLE@Woven by Toyota

DMS: Diffusion-Based Multi-Baseline Stereo Generation for Improving Self-Supervised Depth Estimation

Published in IEEE/CVF International Conference on Computer Vision Workshops (ICCVW 2025), AIM Workshop, 2025

DMS uses a diffusion model to synthesize additional viewpoints along the stereo baseline from unlabeled stereo pairs. These views help recover pixel correspondences in occluded or out-of-frame regions, improving self-supervised stereo matching and monocular depth estimation.

Paper · Code · Supplement · Oral presentation program

DMS generates additional stereo views and uses them to supervise depth estimation with a per-pixel minimum warping loss
DMS pipeline, Figure 1 by Zihua Liu, Yizhou Li, Songyan Zhang, and Masatoshi Okutomi.