Research map / Label-efficient and robust learning / Knowledge distillation
KD-MVS: Knowledge Distillation Based Self-Supervised Learning for Multi-View Stereo
Yikang Ding, Qingtian Zhu, Xiangyue Liu, Wentao Yuan, Haotian Zhang, Chi Zhang
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 internal featuremetric loss to photometric loss (ablation) | teacher model · DTU | Overall (mm): 0.495 -> 0.428 | Improved |
| use internal instead of external ResNet-18 featuremetric loss (ablation) | teacher model · DTU | Overall (mm): 0.459 -> 0.428 | Improved |
| add external ResNet-18 featuremetric loss to photometric loss (ablation) | teacher model · DTU | Overall (mm): 0.495 -> 0.459 | Improved |
| use second round of distillation (student becomes teacher) (ablation) | KD-MVS CasMVSNet · DTU | Overall (mm): 0.361 -> 0.327 | Improved |
| use more than two distillation rounds (ablation) | KD-MVS CasMVSNet · DTU | Overall (mm): 0.327 -> 0.328 | No clear effect |
| add validated mask to GT-depth training (ablation) | student · DTU | Overall (mm): 0.352 -> 0.343 | Improved |
| replace GT depth with validated pseudo depth under l1 loss (ablation) | student · DTU | Overall (mm): 0.343 -> 0.346 | No clear effect |
| replace l1 loss with probabilistic distillation loss (ablation) | student · DTU | Overall (mm): 0.346 -> 0.327 | Improved |
| increase input views from 3 to 5 (ablation) | KD-MVS student · DTU | Overall (mm): 0.341 -> 0.327 | Improved |
| increase input views from 5 to 9 (ablation) | KD-MVS student · DTU | Overall (mm): 0.327 -> 0.331 | No clear effect |
| raise resolution from 512x640 to 864x1152 (ablation) | KD-MVS student · DTU | Overall (mm): 0.359 -> 0.327 | Improved |
| use stricter cross-view check thresholds (tau_conf 0.30, 2.0, 0.020 to 0.15, 1.0, 0.010) (ablation) | KD-MVS student · DTU | Overall (mm): 0.396 -> 0.327 | Improved |
Closest papers by meaning
- Digging Into Uncertainty in Self-Supervised Multi-View Stereo (ICCV 2021 · similarity 0.90)
- Boosting Self-Supervision for Single-View Scene Completion via Knowledge Distillation (CVPR 2024 · similarity 0.87)
- DualNet: Robust Self-Supervised Stereo Matching with Pseudo-Label Supervision (AAAI 2025 · similarity 0.86)
- All-in-One: Transferring Vision Foundation Models into Stereo Matching (AAAI 2025 · similarity 0.86)
- Consistency-aware Self-Training for Iterative-based Stereo Matching (CVPR 2025 · similarity 0.86)
- MVSFormer++: Revealing the Devil in Transformer's Details for Multi-View Stereo (ICLR 2024 · similarity 0.86)
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.
