Research map / Point clouds / Point cloud classification
Image2Point: 3D Point-Cloud Understanding with 2D Image Pretrained Models
Chenfeng Xu, Shijia Yang, Tomer Galanti, Bichen Wu, Xiangyu Yue, Bohan Zhai, Wei Zhan, Peter Vajda, Kurt Keutzer, Masayoshi Tomizuka
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 ImageNet1K pretraining and fully finetune (FIP-ALL) | ResNet152 · ModelNet 3D Warehouse | top-1 accuracy (%): 90.28 -> 91.09 | Improved |
| use a deeper model with image pretraining | FIP-ALL ImageNet1K · ModelNet 3D Warehouse | top-1 accuracy (%): ResNet18 +0.13 vs ResNet152 +0.81 | Improved |
| pretrain on ImageNet21K instead of ImageNet1K | FIP-ALL ResNet50 · ModelNet 3D Warehouse | top-1 accuracy (%): 90.92 -> 91.05 | Improved |
| add image pretraining to ViT | ViT-B-16 / ViT-L-16 · ModelNet 3D Warehouse | top-1 accuracy (%): 84.27 -> 87.77; 83.48 -> 87.66 | Improved |
| add ImageNet1K pretraining | ResNet18 on SemanticKITTI · SemanticKITTI | mIoU (%): 64.75 -> 65.57 | Improved |
| use Cityscapes pretrained HRNetV2-W48 instead of ImageNet1K | HRNetV2-W48 · SemanticKITTI | mIoU (%): +5.25% mIoU over training from scratch (ImageNet1K: 3.41) | Improved |
| pretrain with ImageNet images instead of PointContrast | ModelNet classification · ModelNet 3D Warehouse | top-1 accuracy (%): 89.95 -> 90.88 | Improved |
| add ImageNet1K pretraining in 1-shot setting | ResNet18/50/152 · ModelNet 3D Warehouse 1-shot | top-1 accuracy (%): 26.8 -> 36.8 (+10.0) | Improved |
| add ImageNet1K pretraining as teacher in semi-supervised distillation | ResNet34 · ModelNet 3D Warehouse 1-shot | top-1 accuracy (%): 29.2 -> 41.1 (+11.9) | Improved |
| add ImageNet1K pretraining to speed up training | ResNet50 · ModelNet 3D Warehouse | epochs to 90% top-1: 122 -> 11 epochs (11.1x) | Improved |
| also finetune BN layers in addition to input and output layers (FIP-IO+BN vs FIP-IO) (ablation) | FIP ResNet50 ImageNet1K · ModelNet 3D Warehouse | top-1 accuracy (%): 81.20 -> 89.87 | Improved |
| also update BN mean and variance (IOms vs IO) (ablation) | FIP ResNet50 ImageNet1K · ModelNet 3D Warehouse | top-1 accuracy (%): 81.20 -> 82.94 | Improved |
Closest papers by meaning
- Multi-View Representation is What You Need for Point-Cloud Pre-Training (ICLR 2024 · similarity 0.91)
- Point2Real: Bridging the Gap between Point Cloud and Realistic Image for Open-World 3D Recognition (AAAI 2024 · similarity 0.90)
- Exploring Vision Semantic Prompt for Efficient Point Cloud Understanding (ICML 2025 · similarity 0.90)
- Point Cloud Pre-training with Diffusion Models (CVPR 2024 · similarity 0.89)
- Improved MLP Point Cloud Processing with High-Dimensional Positional Encoding (AAAI 2024 · similarity 0.88)
- UniPre3D: Unified Pre-training of 3D Point Cloud Models with Cross-Modal Gaussian Splatting (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.
