Part-Guided Unsupervised Point Cloud Shape Classification

Table 1. Loss function weight and model accuracy
View table
Table 1. Loss function weight and model accuracy
Model Backbone Contrastive loss /% Reconstruction loss /% Accuracy /%
Rc-Part （Part 2048） PointNet 20 80 88.4
50 50 89.1
80 20 90.2

Table 2. Classfication accuracy on the ModelNet10 and ModelNet40 datasets

View table

Table 2. Classfication accuracy on the ModelNet10 and ModelNet40 datasets

Method	Whether to use labels	Backbone	Accuracy /%
Method	Whether to use labels	Backbone	ModelNet10	ModelNet40
PointNet^［4］	Yes	PointNet	‒	89.2
Jigsaw3D（ModelNet）^［10］	No	PointNet	91.3	87.5
Jigsaw3D（ShapeNet）^［10］	No	PointNet	91.6	87.3
Rotation3D（ShapeNet）^［22］	No	PointNet	‒	88.6
OcCo（ModelNet）^［23］	No	PointNet	91.4	88.7
STRL（ShapeNet）^［24］	No	PointNet	‒	88.3
Multi-view rendering（ModelNet）^［25］	No	PointNet	‒	89.5
ConClu（ModelNet）^［18］	No	PointNet	93.2	89.6
SoftClu（ShapeNet）^［17］	No	PointNet	93.0	88.4
CrossPoint （ShapeNet）^［26］	No	PointNet	‒	89.1
CLR-GAM （ShapeNet）^［12］	No	PointNet	‒	88.9
Rc-Part （ModelNet）	No	PointNet	92.1	89.5
Rc-Part （ShapeNet）	No	PointNet	93.8	89.8
Rc-Part （Part 2048）	No	PointNet	94.0	90.2

Table 3. Classfication accuracy on the ScanObjectNN dataset

View table

Table 3. Classfication accuracy on the ScanObjectNN dataset

Method	Whether to use labels	Backbone	Accuracy /%
Jigsaw3D^［10］	No	PointNet	49.7
OcCo^［23］	No	PointNet	62.1
Multi-view rendering^［25］	No	PointNet	70.2
Rc-Part （Part 2048）	No	PointNet	71.4
Rc-Part （ShapeNet）	No	PointNet	74.0
Rc-Part （ModelNet）	No	PointNet	70.3

Table 4. Accuracy of the Rc-Part network on the ModelNet40 dataset
View table
Table 4. Accuracy of the Rc-Part network on the ModelNet40 dataset
Model Backbone Decoder Accuracy /%
Rc-Part （Part 2048） PointNet FoldingNet 90.2
FC 88.2

Table 5. Ablation experiment results on the ModelNet40 dataset
View table
Table 5. Ablation experiment results on the ModelNet40 dataset
Model Backbone Contrastive learning Reconstruction Accuracy /%
Rc-Part （Part 2048） PointNet ‒ √ 87.2
√ ‒ 83.4
√ √ 90.2

Table 6. Accuracy of different networks on the ModelNet10 and ModelNet40 datasets

View table

Table 6. Accuracy of different networks on the ModelNet10 and ModelNet40 datasets

Method	Whether to use labels	Backbone	Accuracy /%
Method	Whether to use labels	Backbone	ModelNet10	ModelNet40
PointNet++^［5］	Yes	PointNet++	‒	90.7
Jigsaw3D^［10］	No	DGCNN	94.5	90.6
OcCo^［23］	No	DGCNN	92.7	90.2
STRL^［24］	No	DGCNN	‒	90.0
ConClu^［18］	No	DGCNN	94.9	91.8
SoftClu^［17］	No	DGCNN	94.5	91.4
CrossPoint^［26］	No	DGCNN	‒	91.2
CLR-GAM^［12］	No	DGCNN	‒	91.3
Rc-Part （ShapeNet）	No	DGCNN	94.8	91.9
Rc-Part （ModelNet）	No	PointNet++	95.1	92.1

Tools

Get Citation

Copy Citation Text

Haoyang Li, Xie Han, Tingya Liang. Part-Guided Unsupervised Point Cloud Shape Classification[J]. Laser & Optoelectronics Progress, 2025, 62(14): 1439001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites