metadata
base_model:
- Pointcept/PointTransformerV3
datasets:
- labelmaker/arkit_labelmaker
- labelmaker/ALC_pointcept
license: mit
pipeline_tag: image-segmentation
library_name: Pointcept
Models Trained with ARKit LabelMaker
- Please check out out project page labelmaker.org. 😁
- The major dataset is here and the processing code and preprocessed dataset for PTv3 is here.
- Training log tensorboard link
- We made some modification on Pointcept. You can also download per-experiment code base in each experiment folder.
1. 3D Semantic Segmentation
Model: PointTransformer V3 + PPT. Additional dataset: ScanNet++, Structured3D.
| Task | ScanNet | ScanNet200 | ScanNet mIoU | ScanNet200 mIoU | ckpt | config | folder |
|---|---|---|---|---|---|---|---|
| Eval ScanNet on val | train | train + val | 79.1 | - | link | link | link |
| Eval ScanNet200 on val | train + val | train | - | 37.5 | link | link | link |
| Eval on both on test | train + val | train + val | 79.8 | 41.4 | link | link | link |
2. 3D Instance Segmentation
Pre-trained model checkpoints: PointTransformer V3 + PPT. Training dataset: ScanNet/ScanNet200 training split.