Research map / 3D vision / Face and shape
3D-R2N2: A Unified Approach for Single and Multi-view 3D Object Reconstruction
Christopher B. Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, Silvio Savarese
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 |
|---|---|---|---|
| use GRU units instead of LSTM units (1x1x1 conv) (ablation) | 3D-R2N2 · ShapeNet test (5 views) | voxel IoU: 0.499 -> 0.540 | Improved |
| use GRU units instead of LSTM units (3x3x3 conv) (ablation) | 3D-R2N2 · ShapeNet test (5 views) | voxel IoU: 0.539 -> 0.592 | Improved |
| use 3x3x3 convolution (neighboring recurrent connections) instead of 1x1x1 for GRU (ablation) | 3D-R2N2 · ShapeNet test (5 views) | voxel IoU: 0.540 -> 0.592 | Improved |
| use 3x3x3 convolution instead of 1x1x1 for LSTM (ablation) | 3D-R2N2 · ShapeNet test (5 views) | voxel IoU: 0.499 -> 0.539 | Improved |
| replace simple encoder/decoder with deep residual encoder/decoder (ablation) | 3D-GRU-3 · ShapeNet test (5 views) | voxel IoU: 0.592 -> 0.634 | Improved |
| same residual change measured by loss (ablation) | 3D-GRU-3 · ShapeNet test (5 views) | cross-entropy loss: 0.091 -> 0.080 | Improved |
| use Res3D-GRU-3 instead of Kar et al. (which also gets masks and keypoints) | PASCAL VOC single view · PASCAL VOC (PASCAL 3D+), mean | voxel IoU: 0.318 -> 0.571 | Improved |
| use LSTM-1 network instead of Kar et al. | PASCAL VOC single view · PASCAL VOC (PASCAL 3D+), mean | voxel IoU: 0.318 -> 0.456 | Improved |
| increase views from 1 to 5 (car) (ablation) | Res3D-GRU-3 · ShapeNet test, car | voxel IoU: 0.798 -> 0.836 | Improved |
| increase views from 1 to 5 (lamp) (ablation) | Res3D-GRU-3 · ShapeNet test, lamp | voxel IoU: 0.381 -> 0.421 | Improved |
| increase views from 1 to 5 (chair) (ablation) | Res3D-GRU-3 · ShapeNet test, chair | voxel IoU: 0.466 -> 0.550 | Improved |
| use 3D-R2N2 instead of MVS with fewer than 20 views | 3D-R2N2 vs PatchMatch MVS · synthetic CAD models, various texture | voxel IoU: MVS IoU=0 when fewer than 20 views | Improved |
Closest papers by meaning
- Front2Back: Single View 3D Shape Reconstruction via Front to Back Prediction (CVPR 2020 · similarity 0.90)
- Higher-Order Function Networks for Learning Composable 3D Object Representations (ICLR 2020 · similarity 0.89)
- 3D-RCNN: Instance-Level 3D Object Reconstruction via Render-and-Compare (CVPR 2018 · similarity 0.89)
- Pix2Vox: Context-Aware 3D Reconstruction From Single and Multi-View Images (ICCV 2019 · similarity 0.89)
- 3D-PRNN: Generating Shape Primitives With Recurrent Neural Networks (ICCV 2017 · similarity 0.89)
- RayTran: 3D Pose Estimation and Shape Reconstruction of Multiple Objects from Videos with Ray-Traced Transformers (ECCV 2022 · similarity 0.88)
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.
