Multi-Scale Hybrid Feature Extraction and Activation Query Point Cloud Completion Network

Yu Li; Zhenming Yu; Jiawei Deng; Jiaguang Huang; Meini Lü

doi:10.3788/LOP242424

Laser & Optoelectronics Progress, Volume. 62, Issue 16, 1637006(2025)

Multi-Scale Hybrid Feature Extraction and Activation Query Point Cloud Completion Network

Yu Li¹, Zhenming Yu^2、*, Jiawei Deng², Jiaguang Huang², and Meini Lü²

¹School of Information and Communication, Guilin University of Electronic Science and Technology, Guilin 541010, Guangxi , China

²Guangxi Key Laboratory of Machine Vision and Intelligent Control, Wuzhou University, Wuzhou 543002, Guangxi , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(12)

Fig. 1. Point cloud complete network frame diagram

Download full size

Fig. 2. Geometry-aware transformer module

Download full size

Fig. 3. Activate query module

Download full size

Fig. 4. Visualizations of the point cloud completion networks

Download full size

Fig. 5. Enlarged comparison images of partial point cloud network completion networks

Download full size

Fig. 6. Visualization comparison results of ablation experiments. (a) Input; (b) ground truth; (c) our method; (d) FPS+MLP; (e) PP-FD; (f) our method (without AQ)

Download full size

Table 1. Comparison of F-Score values among various networks

View table

Table 1. Comparison of F-Score values among various networks

Category	FoldingNet	PCN	TopNet	PoinTr	SnowFlakeNet	PMPNet++	PointATTN	Ours
Average	0.434	0.584	0.547	0.762	0.742	0.710	0.818	0.839
Airplane	0.710	0.824	0.822	0.923	0.905	0.882	0.818	0.960
Cabinet	0.346	0.480	0.473	0.661	0.625	0.592	0.821	0.724
Car	0.517	0.633	0.581	0.717	0.699	0.607	0.820	0.786
Chair	0.313	0.497	0.441	0.724	0.706	0.676	0.821	0.830
Lamp	0.332	0.522	0.488	0.809	0.790	0.806	0.823	0.898
Sofa	0.279	0.439	0.378	0.644	0.612	0.587	0.811	0.748
Table	0.483	0.633	0.601	0.816	0.808	0.757	0.814	0.894
Watercraft	0.491	0.642	0.589	0.803	0.794	0.770	0.814	0.868

Table 2. Comparison of CD values among various networks

View table

Table 2. Comparison of CD values among various networks

Category	FoldingNet	PCN	TopNet	PoinTr	SnowFlakeNet	PMPNet++	PointATTN	Ours
Average	12.633	11.643	11.204	7.860	8.136	8.363	6.971	6.570
Airplane	8.458	6.887	6.841	4.577	4.904	5.352	6.964	3.625
Cabinet	13.616 11.037	13.354	12.132	9.991	10.333	10.462	6.862	8.916
Car	11.037	10.175	10.277	8.577	8.872	9.597	6.929	7.550
Chair	14.584	13.178	13.078	8.538	8.883	9.063	6.898	6.692
Lamp	14.454	13.515	12.315	7.057	7.265	6.791	6.923	5.353
Sofa	14.972	14.432	13.805	10.050	10.359	10.367	7.116	8.413
Table	13.327	11.142	10.343	7.218	7.466	8.043	7.022	5.947
Watercraft	11.614	10.463	10.751	6.875	7.006	7.231	7.051	5.768

Table 3. Comparison of network results on KITTI dataset
View table
Table 3. Comparison of network results on KITTI dataset
Network F-Score CD
FoldingNet 0.375 14.77
PCN 0.559 11.507
TopNet 0.510 11.385
PoinTr 0.611 10.283
SnowFlakeNet 0.663 9.439
PMPNet++ 0.655 9.027
PointATTN 0.739 8.333
Ours 0.789 7.590

Table 4. F-Score comparison of ablation experiments

View table

Table 4. F-Score comparison of ablation experiments

Category	Ours	FPS+MLP	PP-FD
Average	0.839	0.815	0.819
Airplane	0.960	0.951	0.951
Cabinet	0.724	0.698	0.708
Car	0.786	0.769	0.768
Chair	0.830	0.801	0.813
Lamp	0.898	0.874	0.874
Sofa	0.748	0.706	0.715
Table	0.894	0.872	0.870
Watercraft	0.868	0.848	0.850

Table 5. CD comparison of ablation experiments

View table

Table 5. CD comparison of ablation experiments

Category	Ours	FPS+MLP	PP-FD
Average	6.570	6.953	6.880
Airplane	3.625	3.904	3.953
Cabinet	8.916	9.383	9.072
Car	7.550	7.781	7.820
Chair	6.692	7.371	7.162
Lamp	5.353	5.794	5.747
Sofa	8.413	8.793	8.791
Table	5.947	6.351	6.412
Watercraft	5.768	6.070	6.083

Table 6. Comparison of experimental results of query module
View table
Table 6. Comparison of experimental results of query module
Method CD×10³↓ F-Score↑
Ours （without AQ） 7.467 0.800
Ours 6.570 0.839

Tools

Get Citation

Copy Citation Text

Yu Li, Zhenming Yu, Jiawei Deng, Jiaguang Huang, Meini Lü. Multi-Scale Hybrid Feature Extraction and Activation Query Point Cloud Completion Network[J]. Laser & Optoelectronics Progress, 2025, 62(16): 1637006

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites