Research map / Point clouds / Point cloud classification
Towards More Diverse and Challenging Pre-training for Point Cloud Learning: Self-Supervised Cross Reconstruction with Decoupled Views
Xiangdong Zhang, Shaofeng Zhang, Junchi Yan
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 random crop to cross-reconstruction (ablation) | Point-PQAE · ScanObjectNN PB_T50_RS | accuracy: 87.7 -> 89.6 | Improved |
| add random crop to self-reconstruction (ablation) | Point-MAE · ScanObjectNN PB_T50_RS | accuracy: 88.4 -> 88.8 | Improved |
| add view-relative positional embedding versus none (ablation) | Point-PQAE · ScanObjectNN PB_T50_RS | accuracy: 79.3 -> 89.6 | Improved |
| use relative VRPE instead of absolute positional embedding (ablation) | Point-PQAE · ScanObjectNN PB_T50_RS | accuracy: 87.7 -> 89.6 | Improved |
| use sinusoid VRPE instead of learnable (ablation) | Point-PQAE · ScanObjectNN PB_T50_RS | accuracy: 89.1 -> 89.6 | Improved |
| use rotation instead of jitter augmentation (ablation) | Point-PQAE · ScanObjectNN PB_T50_RS | accuracy: 87.3 -> 89.6 | Improved |
| add normalization and rotation after crop (ablation) | Point-PQAE · ScanObjectNN PB_T50_RS | accuracy: 88.2 -> 89.6 | Improved |
| use Chamfer l2 loss instead of cosine loss (ablation) | Point-PQAE · ScanObjectNN PB_T50_RS | accuracy: 85.2 -> 89.6 | Improved |
| use Chamfer l2 loss instead of Chamfer l1 (ablation) | Point-PQAE · ScanObjectNN PB_T50_RS | accuracy: 89.4 -> 89.6 | Improved |
| add siamese loss (both reconstruction directions) (ablation) | Point-PQAE · ScanObjectNN PB_T50_RS | accuracy: 89.2 -> 89.6 | Improved |
| use positional query in distillation from ReCon versus training from scratch (ablation) | Point-PQAE PQ distillation · ScanObjectNN PB_T50_RS | accuracy: 79.1 -> 88.5 | Improved |
| set minimum crop ratio to 0.6 rather than lower or higher (ablation) | Point-PQAE · ScanObjectNN PB_T50_RS | accuracy | Mixed |
Closest papers by meaning
- CrossPoint: Self-Supervised Cross-Modal Contrastive Learning for 3D Point Cloud Understanding (CVPR 2022 · similarity 0.91)
- Take-A-Photo: 3D-to-2D Generative Pre-training of Point Cloud Models (ICCV 2023 · similarity 0.89)
- From Image Collections to Point Clouds With Self-Supervised Shape and Pose Networks (CVPR 2020 · similarity 0.89)
- Ponder: Point Cloud Pre-training via Neural Rendering (ICCV 2023 · similarity 0.89)
- ToThePoint: Efficient Contrastive Learning of 3D Point Clouds via Recycling (CVPR 2023 · similarity 0.88)
- Sonata: Self-Supervised Learning of Reliable Point Representations (CVPR 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.
