Re-Depth Anything: Test-Time Depth Refinement via Self-Supervised Re-lighting

Accurate depth estimation is fundamental for many computer vision tasks, including 3D reconstruction, robotics, and augmented reality. This paper introduces "Re-Depth Anything," a novel method for test-time depth refinement using self-supervised re-lighting. The approach leverages photometric consistency across various lighting conditions to improve the accuracy and robustness of depth maps inferred from single images or video frames. By re-lighting the scene virtually, the model learns to refine initial depth estimations without requiring additional labeled training data, making it highly practical for real-world applications where ground-truth depth is scarce. This significantly enhances the utility of depth estimation in dynamic and unconstrained environments.