Research map / 3D vision / Depth and stereo
Monocular Scene Reconstruction with 3D SDF Transformers
Weihao Yuan, Xiaodong Gu, Heng Li, Zilong Dong, Siyu Zhu
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 |
|---|---|---|---|
| Replace the 3D CNN U-Net baseline with the SDF transformer (ablation) | ScanNet reconstruction with variance fusion · ScanNet ablation | F-score: 0.647 -> 0.732 | Improved |
| Add variance-based feature fusion (ablation) | 3D CNN U-Net reconstruction baseline · ScanNet ablation | F-score: 0.636 -> 0.647 | Improved |
| Add bottom-level global attention and global context encoding (ablation) | SDF transformer · ScanNet ablation | F-score: 0.732 -> 0.741 | Improved |
| Add post-dilate-attention after the down-up flow (ablation) | SDF transformer with global module · ScanNet ablation | F-score: 0.741 -> 0.754 | Improved |
| Increase sparse attention window size from 1 to 10 (ablation) | SDF transformer · ScanNet ablation | F-score: 0.656 -> 0.754 | Improved |
| Increase sparse attention window size from 8 to 10 (ablation) | SDF transformer · ScanNet ablation | F-score: 0.746 -> 0.754 | Improved |
| Add translation pose noise with σT 1 cm to all frames (ablation) | Ours · ScanNet robustness | F-score: 0.705 -> 0.663 | Hurt |
| Add rotation pose noise with σR 1 deg to all frames (ablation) | Ours · ScanNet robustness | F-score: 0.705 -> 0.351 | Hurt |
| Reduce voxel size from 4cm to 2cm (ablation) | Ours on Scene-709 · Scene-709 | F-score: 0.780 -> 0.832 | Improved |
| Use the proposed 3D SDF transformer instead of the previous best method | ScanNet mesh reconstruction with occlusion-mask evaluation · ScanNet test | F-score: 0.655 -> 0.754 | Improved |
| Use the proposed method instead of VoRTX | ScanNet rendered depth · ScanNet test | Abs Rel: 0.061 -> 0.051 | 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.
- TransformerFusion — TransformerFusion: Monocular RGB Scene Reconstruction using Transformers (NeurIPS 2021) · better on 1 of 1 shared result
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
- Multi-View 3D Reconstruction With Transformers (ICCV 2021 · similarity 0.89)
- 3D Scene Reconstruction from a Single Viewport (ECCV 2020 · similarity 0.89)
- SAF3R: Dynamic Sparse Attention for Feed-Forward 3D Reconstruction Transformers (ECCV 2026 · similarity 0.88)
- TransformerFusion: Monocular RGB Scene Reconstruction using Transformers (NeurIPS 2021 · similarity 0.88)
- Fast Monocular Scene Reconstruction With Global-Sparse Local-Dense Grids (CVPR 2023 · similarity 0.87)
- 3D Reconstruction with Generalizable Neural Fields using Scene Priors (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.
