Reproduce “DepthMaster: Unified Monocular Depth Estimation for Perspective and Panoramic Images” 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

Pengfei Wang, Shihao Wang, Liyi Chen, Zhiyuan Ma, Guowen Zhang, Lei Zhang

Researchers from The Hong Kong Polytechnic University developed DepthMaster, a unified framework for monocular metric depth estimation that accurately predicts depth for both perspective and panoramic images. The model achieved state-of-the-art zero-shot performance across 13 diverse datasets, effectively leveraging perspective priors to address data scarcity in the panoramic domain.

Key insights

  • Unified geometric perception can be achieved by canonicalizing diverse camera inputs into perspective patches and conditioning a standard Transformer with virtual camera parameters, rather than relying on specialized architectures.
  • A feature-level Correspondence Consistency Loss (CCL) is crucial for enforcing global consistency and seamlessly merging depth estimates from overlapping patches, especially for panoramic inputs.
  • Effectively transferring rich geometric priors learned from abundant perspective datasets to the data-scarce panoramic domain is possible through this unified framework, significantly enhancing panoramic depth estimation performance.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.