Research map / Point clouds / Point cloud classification
Take-A-Photo: 3D-to-2D Generative Pre-training of Point Cloud Models
Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, Jiwen Lu
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 TAP 3D-to-2D pre-training | PointMLP · ScanObjectNN PB_T50_RS | accuracy: 87.4 -> 88.5 | Improved |
| add TAP pre-training | DGCNN · ScanObjectNN PB_T50_RS | accuracy: 86.1 -> 86.6 | Improved |
| add TAP pre-training | PointMLP · ModelNet40 | accuracy: 93.7 -> 94.0 | Improved |
| add TAP pre-training | PointMLP · ShapeNetPart | mIoU_I: 86.1 -> 86.9 | Improved |
| pre-train encoder with TAP | 3DETR · ScanNetV2 detection | AP0.5: 37.9 -> 41.4 | Improved |
| use cross-attention that learns projection instead of direct projection (ablation) | TAP photograph module · ScanObjectNN PB_T50_RS | accuracy: 87.6 -> 88.5 | Improved |
| add learnable memory pad token (ablation) | TAP photograph module · ScanObjectNN PB_T50_RS | accuracy: 88.3 -> 88.5 | Improved |
| use formula-derived optical-line queries instead of learnable queries (ablation) | TAP query generator · ScanObjectNN PB_T50_RS | accuracy: 87.5 -> 88.5 | Improved |
| include origin point coordinates in query (ablation) | TAP query generator · ScanObjectNN PB_T50_RS | accuracy: 86.5 -> 88.5 | Improved |
| use 1024-dim cross-attention channels (no bottleneck) (ablation) | TAP photograph module · ScanObjectNN PB_T50_RS | accuracy: 88.5 -> 85.3 | Hurt |
| add VGG perceptual loss to pixel loss (ablation) | TAP objective · ScanObjectNN PB_T50_RS | accuracy: 88.5 -> 87.8 | Hurt |
| use foreground:background loss weight 20:1 (ablation) | TAP objective · ScanObjectNN PB_T50_RS | accuracy: 87.1 -> 88.5 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Point-BERT — Point-BERT: Pre-Training 3D Point Cloud Transformers With Masked Point Modeling (CVPR 2022) · better on 1 of 1 shared result
- Point-M2AE — Point-M2AE: Multi-scale Masked Autoencoders for Hierarchical Point Cloud Pre-training (NeurIPS 2022) · better on 1 of 1 shared result
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
- P2P: Tuning Pre-trained Image Models for Point Cloud Analysis with Point-to-Pixel Prompting (NeurIPS 2022 · similarity 0.90)
- Multi-View Representation is What You Need for Point-Cloud Pre-Training (ICLR 2024 · similarity 0.89)
- Point Cloud Pre-training with Diffusion Models (CVPR 2024 · similarity 0.88)
- Point2Real: Bridging the Gap between Point Cloud and Realistic Image for Open-World 3D Recognition (AAAI 2024 · similarity 0.88)
- GenPC: Zero-shot Point Cloud Completion via 3D Generative Priors (CVPR 2025 · similarity 0.87)
- 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.
