Multi-frame Depth Estimation in Dynamic Scenes

Rui Li Dong Gong Wei Yin Hao Chen Yu Zhu Kaixuan Wang Xiaozhi Chen Jinqiu Sun Yanning Zhang
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

Multi-frame depth estimation that explicitly models dynamic objects and camera motion to overcome the limitations of static-scene assumptions.

Abstract

Multi-frame depth estimation methods typically assume a static scene, which fails when dynamic objects are present. We address this limitation by jointly reasoning about camera motion and per-pixel object motion, enabling reliable depth recovery in dynamic real-world scenes. See the project page for full details, code, and results.