Research map / Point clouds / Point cloud classification
Interpreting Representation Quality of DNNs for 3D Point Cloud Processing
Wen Shen, Qihan Ren, Dongrui Liu, Quanshi Zhang
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 |
|---|---|---|---|
| adversarially train GCNN with rotation and translation attacks (ablation) | GCNN · ModelNet10 | rotation sensitivity: 0.174{ $\pm$0.067} -> 0.034{ $\pm$0.012} | Improved |
| add translation data augmentation (ablation) | PointNet and DGCNN · ModelNet10 | translation sensitivity: PointNet: 0.160{ $\pm$0.067} -> 0.111{ $\pm$ 0.053}; DGCNN: 0.098{ $\pm$0.054} -> 0.048{ $\pm$ 0.024} | Improved |
| add scale data augmentation (ablation) | PointNet and DGCNN · ModelNet10 | scale sensitivity: PointNet: 0.063{ $\pm$0.047} -> 0.025{ $\pm$0.017}; DGCNN: 0.028{ $\pm$0.020} -> 0.020{ $\pm$0.015} | Improved |
| add y-axis rotation augmentation (ablation) | PointNet and DGCNN · ModelNet10 | rotation sensitivity: PointNet: 0.155{ $\pm$0.068} -> 0.108{ $\pm$0.047}; DGCNN: 0.173{ $\pm$0.072} -> 0.158{ $\pm$0.063} | Improved |
| add random-axis rotation augmentation (ablation) | PointNet and DGCNN · ModelNet10 | rotation sensitivity: PointNet: 0.155{ $\pm$0.068} -> 0.057{ $\pm$0.030}; DGCNN: 0.173{ $\pm$0.072} -> 0.052{ $\pm$0.021} | Improved |
| use rotation-based adversarial training rather than random-axis rotation augmentation (ablation) | DGCNN and GCNN · ModelNet10 | rotation sensitivity: DGCNN: 0.052 { $\pm$0.021} -> 0.036{ $\pm$0.015}; GCNN: 0.048{ $\pm$0.024} -> 0.028{ $\pm$0.016} | Improved |
| adversarially train GCNN with rotation and translation attacks (ablation) | GCNN · ModelNet10 | accuracy: 95.1 -> 91.0 | Hurt |
| adversarially train GCNN with rotation and translation attacks (ablation) | GCNN · ShapeNet part | spatial non-smoothness: GCNN: 0.050{ $\pm$0.019} -> adv-GCNN: 0.054{ $\pm$0.110}; excluding knife: 0.049{ $\pm$0.020} -> 0.022{ $\pm$0.008} | Mixed |
| use rotation-invariant architecture (ablation) | PointNet++ and DGCNN · 3D point cloud classification | rotation sensitivity: PointNet++: 0.002{ $\pm$0.001}; DGCNN: 0.010{ $\pm$0.004} | Improved |
| increase selected region size by partitioning into 16 rather than 32 regions (ablation) | PointNet and DGCNN · ModelNet10 | rotation sensitivity: PointNet: 0.159{ $\pm$0.070} -> 0.292{ $\pm$0.117}; DGCNN: 0.174{ $\pm$0.075} -> 0.336{ $\pm$0.144} | Hurt |
| use more training samples (ModelNet40 rather than ModelNet10) (ablation) | PointNet and DGCNN · ModelNet40 | rotation sensitivity: PointNet: 0.159{ $\pm$0.070} -> 0.162{ $\pm$0.073}; DGCNN: 0.174{ $\pm$0.075} -> 0.158{ $\pm$0.065} | No clear effect |
Closest papers by meaning
- Verifiability and Predictability: Interpreting Utilities of Network Architectures for Point Cloud Processing (CVPR 2021 · similarity 0.92)
- Manifold Adversarial Learning for Cross-Domain 3D Shape Representation (ECCV 2022 · similarity 0.89)
- Unsupervised Point Cloud Representation Learning With Deep Neural Networks: A Survey (TPAMI 2023 · similarity 0.88)
- Learning general and distinctive 3D local deep descriptors for point cloud registration (TPAMI 2022 · similarity 0.87)
- Multi-View Representation is What You Need for Point-Cloud Pre-Training (ICLR 2024 · similarity 0.87)
- Learning Representations and Generative Models for 3D Point Clouds (ICML 2018 · similarity 0.87)
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.
