ISBNet: a 3D Point Cloud Instance Segmentation Network with Instance-aware Sampling and Box-aware Dynamic Convolution

Pretrained weights for ISBNet (CVPR 2023).

• Paper: arXiv:2303.00246
• Code: github.com/VinAIResearch/ISBNet

Checkpoints

Full models (use with tools/test.py):

Dataset	AP	AP_50	Config	File
ScanNetV2 val	56.8	73.3	config	scannetv2/head_scannetv2_val.pth
ScanNetV2 val (lightweight)	50.1	68.9	config	scannetv2/head_small_scannetv2_val.pth
ScanNet200 val	24.5	32.7	config	scannet200/head_scannetv2_200_val.pth
S3DIS Area 5	56.3	67.5	config	s3dis/head_s3dis_area5.pth
STPLS3D val	51.2	66.7	config	stpls3d/head_stpls3d_val.pth

Pretrained 3D U-Net backbones (the pretrain: field of each config):

Dataset	File
ScanNetV2 val	scannetv2/pretrain_scannetv2_val.pth
ScanNetV2 train+val (benchmark)	scannetv2/pretrain_scannetv2_trainval.pth
ScanNet200 val	scannet200/pretrain_scannetv2_200_val.pth
S3DIS Area 5	s3dis/pretrain_s3dis_area5.pth
STPLS3D val	stpls3d/pretrain_stpls3d_val.pth

Also included: scannetv2/head_3dyco_scannetv2_val.pth.

Usage

# Download everything into ./pretrains (matches the default `pretrain:` paths in the configs)
hf download TuanNgo/ISBNet --local-dir pretrains

# Evaluate
python3 tools/test.py configs/scannetv2/isbnet_scannetv2.yaml pretrains/scannetv2/head_scannetv2_val.pth

Citation

@inproceedings{ngo2023isbnet,
 author={Tuan Duc Ngo, Binh-Son Hua, Khoi Nguyen},
 booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
 title={ISBNet: a 3D Point Cloud Instance Segmentation Network with Instance-aware Sampling and Box-aware Dynamic Convolution},
 year= {2023}
}