Research map / Point clouds / Point cloud classification
To Supervise or Not to Supervise: Understanding and Addressing the Key Challenges of Point Cloud Transfer Learning
Souhail Hadgi, Lei Li, Maks Ovsjanikov
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 normal-prediction regularization on the first two layers to supervised pre-training | DGCNN supervised pre-training · ModelNet40 | accuracy: 92.78 -> 93.34 | Improved |
| add normal-prediction regularization to supervised pre-training | DGCNN supervised pre-training · ScanObjectNN | accuracy: 84.94 -> 85.91 | Improved |
| add normal-prediction regularization to supervised pre-training | DGCNN supervised pre-training · S3DIS | IoU: 49.07 -> 49.88 | Improved |
| add normal-prediction regularization to contrastive pre-training | DGCNN contrastive pre-training · ModelNet40 | accuracy: 92.78 -> 93.18 | Improved |
| add normal-prediction regularization to contrastive pre-training | DGCNN contrastive pre-training · ScanObjectNN | accuracy: 85.05 -> 85.39 | Improved |
| regularize early layers instead of only the output layer (ablation) | geometric regularization · downstream transfer | downstream performance | Improved |
| use supervised instead of contrastive pre-training under linear probing | point cloud pre-training · ModelNet40, ScanObjectNN | accuracy | Improved |
| use contrastive instead of supervised pre-training under full fine-tuning | DGCNN, MinkowskiNet, PCT · fine-tuning on ModelNet40, ScanObjectNN | accuracy | Improved |
| supervised pre-training before S3DIS fine-tuning | PointNet · S3DIS | IoU: 46.21 -> 37.46 | Hurt |
| contrastive pre-training before S3DIS fine-tuning | MinkowskiNet · S3DIS | IoU: 66.89 -> 60.97 | Hurt |
| contrastive pre-training before S3DIS fine-tuning | PointMLP · S3DIS | IoU: 56.59 -> 58.00 | Improved |
Closest papers by meaning
- Image2Point: 3D Point-Cloud Understanding with 2D Image Pretrained Models (ECCV 2022 · similarity 0.91)
- PointContrast: Unsupervised Pre-training for 3D Point Cloud Understanding (ECCV 2020 · similarity 0.91)
- Multi-View Representation is What You Need for Point-Cloud Pre-Training (ICLR 2024 · similarity 0.90)
- Point Cloud Pre-training with Diffusion Models (CVPR 2024 · similarity 0.89)
- A Unified Approach to Interpreting Self-supervised Pre-training Methods for 3D Point Clouds via Interactions (CVPR 2025 · similarity 0.88)
- UniPre3D: Unified Pre-training of 3D Point Cloud Models with Cross-Modal Gaussian Splatting (CVPR 2025 · 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.
