Reproduce “In Depth We Trust: Reliable Monocular Depth Supervision for Gaussian Splatting” 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
Wenhui Xiao, Ethan Goan, Rodrigo Santa Cruz, David Ahmedt-Aristizabal, Olivier Salvado, Clinton Fookes, Leo Lebrat
The authors introduce a method for using monocular depth to regularize Gaussian Splatting reconstructions selectively. A virtual-stereo consistency check identifies rendered regions with cross-view depth disagreement, where aligned depth is applied, while depth gradients preserve structure. On seven ScanNet++ scenes, the method improves over the no-depth baseline with Marigold, reaching 24.31 dB PSNR and 0.723 δ₁ accuracy in this setting.
Key insights
- Selectively applying absolute depth supervision through a multi-view inconsistency mask is crucial to prevent noisy monocular depth from corrupting robustly reconstructed geometry.
- Leveraging a gradient-based relative depth loss effectively captures fine-grained geometric details from monocular priors, addressing scale ambiguity and avoiding oversmoothed reconstruction.
- Reliably integrating imperfect 2D monocular depth priors can significantly enhance 3D Gaussian Splatting's reconstruction and novel view synthesis, without requiring specialized hardware or compromising established multi-view constraints.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.