Reproduce “GeoSurDepth: Harnessing Foundation Model for Spatial Geometry Consistency-Oriented Self-Supervised Surround-View Depth Estimation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Weimin Liu, Wenjun Wang, Joshua H. Meng
The GeoSurDepth framework introduces a self-supervised approach to surround-view depth estimation by explicitly enforcing spatial geometry consistency, leveraging vision foundation models like DepthAnything V2 for geometric priors and CLIP for feature enhancement. This method achieves state-of-the-art accuracy and produces more edge-aware depth maps on challenging autonomous driving datasets such as DDAD and nuScenes.
Key insights
- Surface normal vectors derived from pseudo scale-ambiguous depth maps from foundation models are inherently invariant to scale and shift ambiguities, making them effective geometric priors for self-supervised learning.
- Explicitly enforcing 3D surface normal consistency and 2D disparity smoothness using foundation model priors significantly improves depth accuracy and edge awareness beyond traditional photometric and smoothness losses.
- Leveraging reconstructed spatial dense depth from multiple views for photometric consistency provides crucial complementary supervision, leading to stronger cross-view geometric coherence and improved view synthesis.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.