ptv3-base.scannet20.pointcept

Model card for ptv3-base.scannet20.pointcept

A Point Transformer V3 point cloud semantic segmentation model (serialized neighborhood attention). Trained on ScanNet (20 classes).

Model Details

• Model Type: Point cloud semantic segmentation
• Model Stats:
  • Params (M): 46.2
  • Input channels: 6
  • Classes: 20
  • Features: 64
• Dataset: ScanNet (20 classes)
• Metrics: mIoU 77.4, OA 92.01 (reference 77.6)
• Paper: Point Transformer V3: Simpler, Faster, Stronger
• Converted from: Pointcept/Pointcept (MIT)
• Library: torch-pointcloud

Install

pip install torch-pointcloud

This checkpoint also needs spconv and flash-attn, which need a build matching your torch and CUDA: see the installation guide.

Usage

import torch
import torch_pointcloud as tp
from torch_pointcloud.utils.data import collate

model, info = tp.create_model(
    "ptv3-base.scannet20.pointcept",
    pretrained=True,
    return_info=True,
)
model = model.cuda().eval()  # GPU-only kernels

# synthetic sample with the keys a dataset provides
num_points = 8192
sample = {
    "pos": torch.randn(num_points, 3),
    "color": torch.rand(num_points, 3) * 255,
    "normal": torch.randn(num_points, 3),
    "segment": torch.zeros(num_points, dtype=torch.long),
    "instance": torch.zeros(num_points, dtype=torch.long),
}
data = info["transform"](sample)
data = collate([data])
data = {key: value.cuda() for key, value in data.items()}

with torch.no_grad():
    logits = model(data.get("x"), data["pos_grid"], data["batch"], pos=data["pos"])

Feature extraction

with torch.no_grad():
    features = model.forward_features(
        data.get("x"),
        data["pos_grid"],
        data["batch"],
        pos=data["pos"],
    )

model.reset_classifier(num_classes=0)
with torch.no_grad():
    features = model(data.get("x"), data["pos_grid"], data["batch"], pos=data["pos"])  # (N, 64)

Citation

@inproceedings{wu2024ptv3,
  title   = {Point Transformer V3: Simpler, Faster, Stronger},
  author  = {Xiaoyang Wu and Li Jiang and Peng-Shuai Wang and Zhijian Liu and Xihui Liu and Yu Qiao and Wanli Ouyang and Tong He and Hengshuang Zhao},
  booktitle = {CVPR},
  year    = {2024}
}

@inproceedings{dai2017scannet,
  title   = {ScanNet: Richly-annotated 3D Reconstructions of Indoor Scenes},
  author  = {Angela Dai and Angel X. Chang and Manolis Savva and Maciej Halber and Thomas Funkhouser and Matthias Nießner},
  booktitle = {CVPR},
  year    = {2017}
}

@software{dujardin2026pytorchpointcloud,
  author  = {Arthur Dujardin},
  title   = {PyTorch PointCloud},
  year    = {2026},
  doi     = {10.5281/zenodo.22159632},
  url     = {https://github.com/arthurdjn/pytorch-pointcloud},
}