Research map / 3D vision / Depth and stereo
Learning Depth Estimation for Transparent and Mirror Surfaces
Alex Costanzino, Pierluigi Zama Ramirez, Matteo Poggi, Fabio Tosi, Stefano Mattoccia, Luigi Di Stefano
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 |
|---|---|---|---|
| in-paint ToM regions with uniform colors to get virtual depth labels (no fine-tuning) | DPT · Booster train (ToM pixels) | MAE (mm): 113.14 -> 57.19 | Improved |
| use N=5 colors with median instead of N=1 (ablation) | DPT virtual depth · Booster train (ToM pixels) | MAE (mm): 65.58 -> 57.19 | Improved |
| fine-tune on virtual depth labels from in-painted images | MiDaS · Booster train (ToM pixels) | MAE (mm): 122.33 -> 93.66 | Improved |
| fine-tune on virtual depth labels from in-painted images | DPT · Booster train (ToM pixels) | MAE (mm): 113.14 -> 70.68 | Improved |
| self-train on own predictions without in-painting (ablation) | MiDaS and DPT · Booster train (ToM pixels) | MAE (mm): 122.33 -> 126.71 (MiDaS); 113.14 -> 118.84 (DPT) | No clear effect |
| in-paint at inference with proxy-segmentation masks instead of fine-tuning (ablation) | DPT · Booster train (All pixels) | delta<1.25: 96.79 -> 93.23 | Hurt |
| use proxy segmentation masks (Trans2Seg/MirrorNet) instead of GT for fine-tuning (ablation) | DPT fine-tuning · Booster train (ToM pixels) | MAE (mm): 67.95 -> 70.67 | Mixed |
| naively extend in-painting to stereo by masking both images (ablation) | RAFT-Stereo · Booster test (ToM pixels) | bad-2 (%): 56.77 -> 57.08 | No clear effect |
| merge stereo disparities with DPT virtual depth on ToM regions | RAFT-Stereo · Booster test (ToM pixels) | bad-2 (%): 56.77 -> 47.54 | Improved |
| merge stereo disparities with DPT virtual depth on ToM regions | CREStereo · Booster test (ToM pixels) | bad-2 (%): 51.83 -> 36.90 | Improved |
| self-train stereo network on own predictions (ablation) | RAFT-Stereo · Booster test (All pixels) | bad-2 (%): 17.42 -> 18.42 | Hurt |
| use proxy masks instead of GT masks for merged stereo labels (ablation) | RAFT-Stereo · Booster test (All pixels) | bad-2 (%): 17.42 -> 19.56 | Mixed |
Closest papers by meaning
- Mirror3D: Depth Refinement for Mirror Surfaces (CVPR 2021 · similarity 0.89)
- Booster: A Benchmark for Depth From Images of Specular and Transparent Surfaces (TPAMI 2023 · similarity 0.88)
- 3D Distillation: Improving Self-Supervised Monocular Depth Estimation on Reflective Surfaces (ICCV 2023 · similarity 0.87)
- Learning Transparent Object Matting (IJCV 2019 · similarity 0.87)
- Seeing and Seeing Through the Glass: Real and Synthetic Data for Multi-Layer Depth Estimation (ICCV 2025 · similarity 0.87)
- A Simple yet Universal Framework for Depth Completion (NeurIPS 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.
