- 2023.5.18 We found that it was unreasonable to use cross product to detect occlusion in manhattan post-processing layout.py#L70. We fixed this problem here. In addition, we add some optimizations to force the layout coordinates to align to satisfy the constraints based on the Manhattan assumption.
- Simplify polygon by DP algorithm
- Detect occlusion, calculating box fill with 1
- Fill in reasonable sampling section
- Output processed polygon
It works, and a performance comparison on the MatterportLayout dataset:
| Method | 2D IoU(%) | 3D IoU(%) | RMSE | |
|---|---|---|---|---|
| without post-proc | 83.52 | 81.11 | 0.204 | 0.951 |
| original post-proc | 83.12 | 80.71 | 0.230 | 0.936 |
| optimized post-proc | 83.48 | 81.08 | 0.214 | 0.940 |
original:
optimized:
