Research map / 3D vision / Depth and stereo
"Smoothness, Synthesis, and Sampling: Re-thinking Unsupervised Multi-View Stereo with DIV Loss"
Alex Rich, Noah Stier, Pradeep Sen, Tobias Hollerer
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add clamped 2nd-order depth smoothness (ablation) | DIV-MVS baseline loss · DTU | Overall (mm): 0.361 -> 0.346 | Improved |
| clamp the 1st-order smoothness gradient (ablation) | DIV-MVS baseline loss · DTU | Overall (mm): 0.361 -> 0.366 | Hurt |
| use 2nd-order smoothness without clamping (ablation) | DIV-MVS baseline loss · DTU | Overall (mm): 0.361 -> 0.365 | Hurt |
| add learned reference-view synthesis (no occlusion masks) (ablation) | DIV-MVS with clamped 2nd-order · DTU | Overall (mm): 0.346 -> 0.339 | Improved |
| add occlusion masking without learned synthesis (ablation) | DIV-MVS with clamped 2nd-order · DTU | Overall (mm): 0.346 -> 0.351 | Hurt |
| add occlusion masking to learned synthesis (ablation) | DIV-MVS · DTU | Overall (mm): 0.346 -> 0.335 | Improved |
| add supervision-view sampling beyond MVS input views (ablation) | DIV-MVS · DTU | Overall (mm): 0.335 -> 0.330 | Improved |
| replace pipeline loss with DIV loss | DIV-RC (RC-MVSNet) · DTU | Overall (mm): 0.350 -> 0.333 | Improved |
| replace pipeline loss with DIV loss | DIV-CL (CL-MVSNet) · DTU | Overall (mm): 0.330 -> 0.321 | Improved |
| replace pipeline loss with DIV loss | DIV-MVS · DTU | Abs. depth error (mm): 19.34 -> 16.32 | Improved |
| replace pipeline loss with DIV loss | DIV-MVS · DTU | Training memory (GB): 10.50 -> 10.52 | No clear effect |
| train DIV-MVS on DTU only and test zero-shot on Tanks and Temples intermediate | DIV-MVS · Tanks and Temples intermediate | F-score: 37.03 -> 38.35 (vs CL-MVSNet) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- RC-MVSNet — RC-MVSNet: Unsupervised Multi-View Stereo with Neural Rendering (ECCV 2022) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Rethinking Depth Estimation for Multi-View Stereo: A Unified Representation (CVPR 2022 · similarity 0.92)
- RC-MVSNet: Unsupervised Multi-View Stereo with Neural Rendering (ECCV 2022 · similarity 0.91)
- CL-MVSNet: Unsupervised Multi-View Stereo with Dual-Level Contrastive Learning (ICCV 2023 · similarity 0.91)
- Just a Few Points Are All You Need for Multi-View Stereo: A Novel Semi-Supervised Learning Method for Multi-View Stereo (ICCV 2021 · similarity 0.90)
- MVSFormer++: Revealing the Devil in Transformer's Details for Multi-View Stereo (ICLR 2024 · similarity 0.89)
- A Global Depth-Range-Free Multi-View Stereo Transformer Network with Pose Embedding (NeurIPS 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
