Models¶

torch-pointcloud ships 36 architectures across point cloud classification, semantic and instance segmentation, object detection, self-supervised pretraining, and generative modeling. Every model is registered with a timm-style factory, and make it switch between tasks or backbones and reset the head for downstream tasks (e.g. features extraction).

import torch_pointcloud as tp

model = tp.create_model(
    "pointnext-sm.scanobjectnn.openpoints", 
    task="classification",
    pretrained=True,
)

Browse available checkpoints with:

import torch_pointcloud as tp

models = tp.list_models(task="classification")
# or task="segmentation", "detection", ...

Five pretrained checkpoints on two committed sample clouds: object classification, part segmentation, scene segmentation, 3D detection, and LiDAR segmentation

Tasks¶

Classification

One label per cloud: run, evaluate and fine-tune a classifier.
Semantic segmentation

One label per point: voxelization, full-resolution predictions, mIoU.
Part segmentation

Category-conditioned part labels and the ShapeNetPart protocol.
Object detection

Oriented boxes: decode, filter with NMS, score with mAP.
Feature maps

The representation under the head: embeddings, retrieval, PCA.

Classification¶

Best for shape classification (ModelNet40, ScanObjectNN, ShapeNet objects). Inputs are single object scans, outputs are scene-level class predictions.

Model	Paper	Benchmark
PointNet	PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation	–
PointNet++	PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space	ModelNet40 OA: 92.67 / 92.8
DGCNN	Dynamic Graph CNN for Learning on Point Clouds	ModelNet40 OA: 92.46 / 93.6
PointCNN	PointCNN: Convolution On \(\mathcal{X}\)-Transformed Points	–
PointConv	PointConv: Deep Convolutional Networks on 3D Point Clouds	ModelNet40 OA: 92.02 / 92.5
PointMLP	Rethinking Network Design and Local Geometry in Point Cloud: A Simple Residual MLP Framework	ModelNet40 OA: – / 94.1
PointNeXt	PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies	ModelNet40 OA: 92.1 / 94.0
Point Transformer V1	Point Transformer	–
Point Transformer V2	Point Transformer V2: Grouped Vector Attention and Partition-based Pooling	–
Point Transformer V3	Point Transformer V3: Simpler, Faster, Stronger	–
PVCNN	Point-Voxel CNN for Efficient 3D Deep Learning	–
Point-Mamba	PointMamba: A Simple State Space Model for Point Cloud Analysis	ModelNet40 OA: 93.64 / 93.6
OctFormer	OctFormer: Octree-based Transformers for 3D Point Clouds	ModelNet40 OA: 89.02 / 92.7

Segmentation¶

Best for dense per-point labeling (S3DIS, ScanNet, SemanticKITTI). Inputs are large scenes, outputs are per-point class predictions.

Model	Paper	Benchmark
PointNet++	PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space	S3DIS-A5 mIoU: 63.59 / 63.6
DGCNN	Dynamic Graph CNN for Learning on Point Clouds	ScanNet20 mIoU: 50.58 / 49.6
KPConv	KPConv: Flexible and Deformable Convolution for Point Clouds	S3DIS-A5 mIoU: 65.66 / 67.3
PointNeXt	PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies	S3DIS-A5 mIoU: 63.01 / 63.4
Point Transformer V3	Point Transformer V3: Simpler, Faster, Stronger	ScanNet20 mIoU: 76.04 / 77.6
RandLA-Net	RandLA-Net: Efficient Semantic Segmentation of Large-Scale Point Clouds	SemanticKITTI mIoU: – / 53.1
SPVCNN	Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution	SemanticKITTI mIoU: 62.4 / 63.8
SPUNet	4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks	ScanNet20 mIoU: 70.02 / 75.67
OctFormer	OctFormer: Octree-based Transformers for 3D Point Clouds	ScanNet20 mIoU: 74.78 / 74.8
SphereFormer	Spherical Transformer for LiDAR-based 3D Recognition	SemanticKITTI mIoU: – / 67.8
PVCNN / PVCNN++	Point-Voxel CNN for Efficient 3D Deep Learning	S3DIS-A5 mIoU: 57.51 / 56.64

Instance segmentation¶

Predict per-point instance masks on top of semantics (ScanNet, S3DIS).

Model	Paper	Benchmark
SPFormer-UNet	Superpoint Transformer for 3D Scene Instance Segmentation	–

Detection¶

Predict 3D bounding boxes for indoor scenes (ScanNet, SUN RGB-D) or driving scenes (KITTI, nuScenes, Waymo).

Model	Paper	Benchmark
VoteNet	Deep Hough Voting for 3D Object Detection in Point Clouds	ScanNet mAP@25: 58.35 / 58.6
3DETR	An End-to-End Transformer Model for 3D Object Detection	SunRGBD mAP@25: 58.20 / 58.0
PointPillars	PointPillars: Fast Encoders for Object Detection from Point Clouds	KITTI mod. mAP: 62.86 / 64.08
SECOND	SECOND: Sparsely Embedded Convolutional Detection	KITTI mod. mAP: 66.11 / 66.25
PointRCNN	PointRCNN: 3D Object Proposal Generation and Detection from Point Cloud	KITTI mod. mAP: 63.56 / 68.41
VoxelNeXt	VoxelNeXt: Fully Sparse VoxelNet for 3D Object Detection and Tracking	nuScenes mAP: – / 60.5
Voxel-Mamba	Voxel Mamba: Group-Free State Space Models for Point Cloud based 3D Object Detection	–
LION	LION: Linear Group RNN for 3D Object Detection in Point Clouds	nuScenes mAP: – / 68.0

Self-supervised pretraining¶

Backbones pretrained without labels, registered with task="base"; the fine-tuned classification / segmentation heads are registered under their downstream task.

Model	Paper	Benchmark
Point-MAE	Masked Autoencoders for Point Cloud Self-supervised Learning	ModelNet40 OA: 93.35 / 94.04
Point-BERT	Point-BERT: Pre-training 3D Point Cloud Transformers with Masked Point Modeling	ModelNet40 OA: 93.07 / 93.19
Point-M2AE	Point-M2AE: Multi-scale Masked Autoencoders for Hierarchical Point Cloud Pre-training	ModelNet40 OA: 92.87 / 93.43
PointGPT	PointGPT: Auto-regressively Generative Pre-training from Point Clouds	ModelNet40 OA: 94.37 / 94.4
PointMamba	PointMamba: A Simple State Space Model for Point Cloud Analysis	ModelNet40 OA: 93.64 / 93.6
Sonata	Sonata: Self-Supervised Learning of Reliable Point Representations	ScanNet20 lin. mIoU: 71.93 / 72.5
Concerto	Concerto: Joint 2D-3D Self-Supervised Learning Emerges Spatial Representations	ScanNet20 lin. mIoU: 77.68 / –
Utonia	Utonia: Toward One Encoder for All Point Clouds	ScanNet20 lin. mIoU: 71.11 / –