Semantic Segmentation for Real Point Cloud Scenes via Bilateral Augmentation and Adaptive Fusion (CVPR 2021)
- My personal experimental implementation. Since the original Repo missing this part.
- SemanticKITTI part of the code borrowed from RandLA-Net, and almost identical.
- Mostly same as original Repo.
Please refer to the two repo mentioned above.
On the validation set, the results are close to those reported by the original Repo.
| Model | mIoU | car | bicycle | motorcycle | truck | other-vehicle | person | bicyclist | motorcyclist | road | parking | sidewalk | other-ground | building | fence | vegetation | trunk | terrain | pole | traffic-sign |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Original Repo Report | 58.71 | 94.75 | 12.41 | 52.01 | 80.57 | 46.41 | 59.77 | 75.02 | 0.01 | 93.11 | 44.72 | 80.74 | 2.56 | 89.84 | 51.45 | 87.09 | 66.55 | 77.33 | 57.92 | 43.23 |
| Ours | 58.83 | 94.54 | 11.29 | 50.35 | 87.36 | 52.44 | 59.05 | 73.08 | 0.00 | 93.24 | 46.84 | 81.05 | 1.31 | 89.33 | 48.42 | 86.30 | 66.49 | 77.44 | 57.30 | 42.02 |
| Log | Pretrain-model | Our Validation results | Original Repo Validation results |
|---|---|---|---|
| Link | Link | Link | Link |