Figure 1.
Cost Volume Regularization Network Architecture
Figure 2.
Self-attention Architecture
Figure 3.
The Enhanced Feature Extraction Block
Figure 4.
Residual Block
Figure 5.
Pre-activated Residual Block
Figure 6.
Improved Cost Volume Regularization Network
Figure 7.
Partial Scenes in DTU Dataset
TABLE I.
Experimental Parameters
| parameter | value |
|---|---|
| batch size | 1 |
| learning rate | 0.001 |
| epoch | 16 |
| Adam-β1 | 0.9 |
| Adamβ2 | 0.999 |
Figure 8.
Loss Curve
Figure 9.
Depth Map Results
TABLE II.
Comparison of Different 3D Reconstruction
| Methods | (Acc)/mm | (Comp)/mm | (OA)/mm |
|---|---|---|---|
| Colmap | 0.400 | 0.664 | 0.532 |
| Gipuma | 0.283 | 0.873 | 0.578 |
| Camp | 0.835 | 0.554 | 0.695 |
| MVSNet | 0.456 | 0.574 | 0.515 |
| PointMVSNet | 0.435 | 0.475 | 0.455 |
| SelfResMVSNet | 0.442 | 0.562 | 0.502 |
Figure 10.
Comparison of the generated point cloud models