Research map / Point clouds / Point cloud classification
CO-Net: Learning Multiple Point Cloud Tasks at Once with A Cohesive Network
Tao Xie, Ke Wang, Siyi Lu, Yukun Zhang, Kun Dai, Xiaoyu Li, Jie Xu, Li Wang, Lijun Zhao, Xinyu Zhang, Ruifeng Li
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 searched subnet with random subnet under parameter constraint (ablation) | CO-Net (base) · ScanObjectNN / S3DIS / SUN RGBD | OA / mIoU / mAP: cls 87.3 -> 88.6, seg 63.6 -> 64.8, det 62.1 -> 62.9 | Improved |
| let each task have task-specific parameters instead of sharing all backbone parameters (ablation) | CO-Net (base) · three tasks | S-Score: 210.1 -> 216.1 | Improved |
| make normalization layers task-specific as well as conv layers (ablation) | CO-Net (base) · three tasks | S-Score: 214.4 -> 216.1 (EXP2 vs EXP3) | No clear effect |
| add sign-based gradient surgery (ablation) | CO-Net (base) · three tasks | S-Score: 213.2 -> 216.1 | Improved |
| use sign-based gradient surgery instead of PCGrad (ablation) | CO-Net (base) · three tasks | S-Score: 215.1 -> 216.1 | Improved |
| use sign-based gradient surgery instead of DWA (ablation) | CO-Net (base) · three tasks | S-Score: 214.2 -> 216.1 | Improved |
| joint training and NAS versus individually trained searched architecture (ablation) | CO-Net (base) · ScanObjectNN / S3DIS / SUN RGBD | OA / mIoU / mAP: cls 88.6 -> 88.4, seg 64.8 -> 65.4, det 62.9 -> 62.6; params 4.300 -> 3.362 | Mixed |
| set lambda to 0.5 versus 0.25 and 0.75 (ablation) | CO-Net (base) · three tasks | S-Score: 215.4 / 215.5 -> 216.1 | Improved |
| freeze task-shared parameters and train only task-specific ones for a new task | CO-Net (base) · ModelNet40 incremental | overall accuracy: 90.7 -> 92.1 with 62.15% extra parameters | Improved |
| joint multi-task CO-Net versus segmentation-specialized models | CO-Net · S3DIS Area 5 | mIoU / OA: CO-Net (large) 68.0 vs Point Transformer 70.4 | No clear effect |
Closest papers by meaning
- CO-Net++: A Cohesive Network for Multiple Point Cloud Tasks at Once With Two-Stage Feature Rectification (TPAMI 2024 · similarity 0.94)
- Poly-PC: A Polyhedral Network for Multiple Point Cloud Tasks at Once (CVPR 2023 · similarity 0.89)
- A Survey of Label-Efficient Deep Learning for 3D Point Clouds (TPAMI 2024 · similarity 0.87)
- DeepLA-Net: Very Deep Local Aggregation Networks for Point Cloud Analysis (CVPR 2025 · similarity 0.87)
- ASSANet: An Anisotropic Separable Set Abstraction for Efficient Point Cloud Representation Learning (NeurIPS 2021 · similarity 0.87)
- LinNet: Linear Network for Efficient Point Cloud Representation Learning (NeurIPS 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.
