Multi-View 3D Reconstruction Method Based on Self-Attention Mechanism

Fig. 5. Relationship between the overall error and GPU memory and running time, the image size is 1152×864. (a) Relationship between overall error and GPU memory consumption; (b) relationship between overall error and running time

Download full size

Fig. 6. Depth map comparison of ablation experiment. (a) Depth map without improvement; (b) improved depth map; (c) GT

Download full size

Fig. 7. Point cloud local magnification comparison map in the ablation experiment

Download full size

Table 1. Test results of different methods on the DTU dataset

View table

Table 1. Test results of different methods on the DTU dataset

Method	Acc /mm	Comp /mm	Overall /mm
Camp	0.835	0.554	0.695
Furu	0.613	0.941	0.777
Tola	0.342	1.190	0.766
Gipuma	0.283	0.873	0.578
MVSNet	0.396	0.527	0.462
R-MVSNet	0.383	0.452	0.417
P-MVSNet	0.406	0.434	0.420
Point-MVSNet	0.342	0.411	0.376
Fast-MVSNet	0.336	0.403	0.370
CasMVSNet	0.325	0.385	0.355
CVP-MVSNet	0.296	0.406	0.351
M3VSNet	0.636	0.531	0.583
PatchmatchNet	0.427	0.277	0.352
Proposed method	0.427	0.261	0.344

Table 2. Results of different methods on Tanks and Temples dataset
View table
Table 2. Results of different methods on Tanks and Temples dataset
Method F-score
COLMAP 27.24
R-MVSNet 24.91
CasMVSNet 31.12
PatchmatchNet 32.31
Proposed method 32.72

Table 3. Comparison of quantitative results of ablation experiment
View table
Table 3. Comparison of quantitative results of ablation experiment
Module Acc /mm Comp /mm Overall /mm GPU /GB Running time /s Vertices
MsFe 0.427 0.277 0.352 10.89 0.210 51527707
SA-MsFe 0.427 0.261 0.344 7.8 0.562 52289896

Tools

Get Citation

Copy Citation Text

Guangzhao Zhu, Bo Wei, Afeng Yang, Xin Xu. Multi-View 3D Reconstruction Method Based on Self-Attention Mechanism[J]. Laser & Optoelectronics Progress, 2023, 60(16): 1615009

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites