Reproduce “MiDaS v3.1 -- A Model Zoo for Robust Monocular Relative 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
Reiner Birkl, Diana Wofk, Matthias Müller
Intel Labs' MiDaS v3.1 improves monocular relative depth estimation by integrating a "model zoo" of state-of-the-art vision transformer architectures. This collection of models offers various performance-runtime trade-offs, with the top-performing BEiT512-L model achieving up to a 36% reduction in average relative zero-shot error compared to its predecessor.
Key insights
- Transformer-based architectures, particularly BEiT and Swin variants, consistently outperform advanced convolutional networks for monocular relative depth estimation within the MiDaS framework.
- The systematic integration of diverse backbones enables the creation of a "model zoo" that offers various performance-runtime trade-offs, catering to different application needs.
- Expanding the training dataset mix to include datasets like KITTI and NYU Depth v2 significantly improves overall model quality and generalizability, even if it affects strict zero-shot behavior on those specific datasets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.