Research map / 3D vision / Face and shape
3D Scene Reconstruction With Multi-Layer Depth and Epipolar Transformers
Daeyun Shin, Zhile Ren, Erik B. Sudderth, Charless C. Fowlkes
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 overhead EFT height map to first-layer depth D1 (ablation) | multi-layer depth model · SUNCG objects (5cm) | precision / recall: 0.525 -> 0.553 precision, 0.212 -> 0.275 recall | Improved |
| add overhead EFT height map to four depth layers (ablation) | multi-layer depth model · SUNCG objects (5cm) | precision / recall: 0.499 -> 0.519 precision, 0.417 -> 0.457 recall | Improved |
| add depth layers 2-4 to the first layer (ablation) | multi-layer depth model · SUNCG objects (5cm) | precision / recall: 0.525 -> 0.499 precision, 0.212 -> 0.417 recall | Mixed |
| add semantic segmentation network features to EFT input (ablation) | overhead height map network · SUNCG overhead height map | L1 error: 0.141 -> 0.132 | Improved |
| add depth network CNN features rather than only RGB, mask and best guess depth (ablation) | overhead height map network · SUNCG overhead height map | L1 error: 0.144 -> 0.132 | Improved |
| add back surfaces and more layers (D1..5 vs D1) in ground-truth coverage (ablation) | multi-layer depth representation · SUNCG ground-truth coverage (5cm) | recall: 0.237 -> 0.924 | Improved |
| add overhead virtual view to D1..5 ground-truth layers (ablation) | multi-layer depth representation · SUNCG ground-truth coverage (5cm) | recall: 0.924 -> 0.932 | Improved |
| include back surfaces (layers 1-5 vs 1,3,5 only) in ground truth layers (ablation) | multi-layer depth representation · SUNCG scene frustum coverage | coverage: 82% -> 92% | Improved |
| viewer-centered multi-layer + EFT instead of object-based Factored3D | scene reconstruction · ScanNet (1000 images, 10cm) | precision / recall: 0.132 -> 0.221 precision, 0.191 -> 0.358 recall | Improved |
| viewer-centered multi-layer + EFT instead of object-based Factored3D | scene reconstruction · SUNCG test objects (voxel IoU) | voxel IoU: 0.102 to 0.310 | Improved |
Closest papers by meaning
- TransformerFusion: Monocular RGB Scene Reconstruction using Transformers (NeurIPS 2021 · similarity 0.90)
- Multi-view Depth Estimation using Epipolar Spatio-Temporal Networks (CVPR 2021 · similarity 0.88)
- MVSTER: Epipolar Transformer for Efficient Multi-View Stereo (ECCV 2022 · similarity 0.88)
- Complete 3D Scene Parsing from an RGBD Image (IJCV 2018 · similarity 0.88)
- Epipolar Transformers (CVPR 2020 · similarity 0.88)
- Know Your Neighbors: Improving Single-View Reconstruction via Spatial Vision-Language Reasoning (CVPR 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.
