Research map / 3D vision / Face and shape
Mesh R-CNN
Georgia Gkioxari, Jitendra Malik, Justin Johnson
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 voxel branch with cubify and mesh refinement instead of sphere-initialized refinement (ablation) | Mesh R-CNN · Pix3D S1 | APmesh: 40.5 -> 55.4 | Improved |
| add voxel branch (arbitrary topology) vs sphere-init on objects with holes (ablation) | Mesh R-CNN · ShapeNet holes test set (Pretty) | Chamfer: 0.196 -> 0.171 | Improved |
| add voxel branch vs sphere-init on the full ShapeNet test set (ablation) | Mesh R-CNN · ShapeNet full test set (Best) | Chamfer: 0.132 -> 0.133 | No clear effect |
| add mesh refinement branch to voxel predictions (ablation) | Voxel-Only · Pix3D S1 | APmesh: 6.8 -> 55.4 | Improved |
| add mesh subdivision (Pixel2Mesh+ vs Sphere-Init) (ablation) | sphere-based mesh refinement · ShapeNet full test set (Best) | Chamfer: 0.132 -> 0.132 | No clear effect |
| use deeper residual refinement branch instead of lighter nonresidual one (ablation) | Mesh R-CNN · ShapeNet full test set (Best) | Chamfer: 0.133 -> 0.133 | No clear effect |
| add edge length regularizer (Pretty vs Best) (ablation) | Mesh R-CNN · ShapeNet (protocol of Pixel2Mesh) | Chamfer: 0.306 -> 0.391 | Hurt |
| initialize from COCO instead of ImageNet (ablation) | Mesh R-CNN · Pix3D S1 | APmesh: 52.9 -> 55.4 | Improved |
| use three refinement stages instead of one (ablation) | Mesh R-CNN · Pix3D S1 | APmesh: 52.4 -> 55.4 | Improved |
| use three refinement stages instead of two (ablation) | Mesh R-CNN · Pix3D S1 | APmesh: 54.5 -> 55.4 | Improved |
| use Mesh R-CNN instead of Sphere-Init on unseen shapes (ablation) | Mesh R-CNN · Pix3D S2 (disjoint models) | APmesh: 24.8 -> 28.7 | Improved |
| evaluate on ground-truth regions instead of detected boxes | Mesh R-CNN · Pix3D S1 | APmesh: 55.4 -> 49.1 | Hurt |
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.
- Pixel2Mesh+ — Pixel2Mesh: Generating 3D Mesh Models from Single RGB Images (ECCV 2018) · better on 5 of 6 shared results
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
- Pixel2Mesh: 3D Mesh Model Generation via Image Guided Deformation (TPAMI 2020 · similarity 0.90)
- Primal-Dual Mesh Convolutional Neural Networks (NeurIPS 2020 · similarity 0.90)
- Survey and Evaluation of Neural 3D Shape Classification Approaches (TPAMI 2021 · similarity 0.89)
- FeaStNet: Feature-Steered Graph Convolutions for 3D Shape Analysis (CVPR 2018 · similarity 0.89)
- Voxel R-CNN: Towards High Performance Voxel-based 3D Object Detection (AAAI 2021 · similarity 0.88)
- Learning Unsupervised Hierarchical Part Decomposition of 3D Objects From a Single RGB Image (CVPR 2020 · 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.
