Research map / Point clouds / Point cloud classification
ToThePoint: Efficient Contrastive Learning of 3D Point Clouds via Recycling
Xinglin Li, Jiajing Chen, Jinhui Ouyang, Hanhui Deng, Senem Velipasalar, Di Wu
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 |
|---|---|---|---|
| Recycle discarded point features and align them with max-pooled features (ablation) | ToThePoint · ScanObjectNN | SVM accuracy: 61.45% -> 74.70% (PointNet); 80.38% -> 81.93% (DGCNN) | Improved |
| Use two branches instead of one, retaining recycling (ablation) | ToThePoint · ScanObjectNN | SVM accuracy: 7.06%; 8.27% accuracy reduction with one branch (PointNet; DGCNN) | Improved |
| Use two branches instead of one, retaining recycling (ablation) | ToThePoint · ModelNet40 | SVM accuracy: 2.76%; 0.41% accuracy reduction with one branch (PointNet; DGCNN) | Improved |
| Add recycling to the two-branch model (ablation) | ToThePoint · ModelNet40 | SVM accuracy: 6.89%; 0.37% accuracy reduction without recycling (PointNet; DGCNN) | Improved |
| Add recycling to the two-branch model (ablation) | ToThePoint · ScanObjectNN | SVM accuracy: 13.25%; 1.56% accuracy reduction without recycling (PointNet; DGCNN) | Improved |
| Use ToThePoint pretraining rather than CrossPoint | DGCNN · ModelNet40-C (LiDAR, severity 1) | accuracy: 83.69±0.29 -> 83.80±0.32 | Improved |
| Use ToThePoint pretraining rather than CrossPoint | PointNet · ModelNet40-C (LiDAR, severity 1) | accuracy: 81.12±0.44 -> 80.97±0.27 | Hurt |
| Use ToThePoint pretraining rather than the strongest listed baseline | PointNet and DGCNN · ScanObjectNN | SVM accuracy: 73.67 -> 74.70 (PointNet); 81.43 -> 81.93 (DGCNN) | Improved |
| Use ToThePoint pretraining rather than CrossPoint | DGCNN · ModelNet40, 5-way 10-shot | accuracy: 91.12 ± 4.93 -> 92.73 ± 4.79 | Improved |
| Use ToThePoint pretraining rather than CrossPoint | DGCNN · ModelNet40, 5-way 20-shot | accuracy: 94.56 ± 3.23 -> 95.10 ± 2.95 | Improved |
| Use ToThePoint pretraining rather than CrossPoint | DGCNN · ScanObjectNN, 5-way 10-shot | accuracy: 77.24 ±7.13 -> 78.13 ± 7.29 | Improved |
| Use ToThePoint pretraining rather than CrossPoint | PointNet · ScanObjectNN, 10-way 10-shot | accuracy: 66.61 ±3.96 -> 66.71 ± 5.40 | No clear effect |
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.
- CrossPoint — CrossPoint: Self-Supervised Cross-Modal Contrastive Learning for 3D Point Cloud Understanding (CVPR 2022) · better on 13 of 13 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
- CrossPoint: Self-Supervised Cross-Modal Contrastive Learning for 3D Point Cloud Understanding (CVPR 2022 · similarity 0.92)
- Complete Structure Guided Point Cloud Completion via Cluster- and Instance-Level Contrastive Learning (NeurIPS 2025 · similarity 0.92)
- Sonata: Self-Supervised Learning of Reliable Point Representations (CVPR 2025 · similarity 0.90)
- Point Contrastive Prediction with Semantic Clustering for Self-Supervised Learning on Point Cloud Videos (ICCV 2023 · similarity 0.90)
- Learning Transferable Features for Point Cloud Detection via 3D Contrastive Co-training (NeurIPS 2021 · similarity 0.89)
- LPCG: A Self-conditional Architecture for Labeled Point Cloud Generation (AAAI 2025 · 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.
