This repository is merged with https://github.com/jfzhang95/pytorch-deeplab-xception and is success of https://github.com/sunggukcha/deeplabv3plus-bdd100k-drivablearea. Please follow installation policies of the repositories above.
For BDD100k/drivable_area semantic segmentation, I added
- bdd100k drivable area dataloader, and training/val/test scripts.
- prediction visualization for both color (visual result) and id (greyscale png file for submission).
- added Group Noramlization.
- deeplabv3 which is without deeplabv3+ decoder, but with aspp only.
- WRN as backbone is added (original code from mapillary@github)
- additional visualization that marks corrects, missed and wrong pixels.
- IBN-Net by github.com/XingangPan/IBN-Net/
- EfficientNet added which is implemented by https://github.com/lukemelas/EfficientNet-PyTorch.
- Feature Pyramid Networks(FPNs) for semantic segmentation added (version: Panoptic Feature Pyramid Networks).
For more detail, please visit the repositories above.
Single 12GB GPU
| Backbone | Normalization | mIoU in test |
|---|---|---|
| ResNet50 | Group-16 | 85.00% |
| ResNet101 | IGN-a-16 | 85.12% |
| ResNet101 | Group-16 | 85.33% |
| ResNet152 | Group-16 | 85.45% |
IGN-a-16 denotes instance group normalization with channel-grouping number 16, replacing BN of IBNNet-a with GN16. Group-16 denotes group normalization with channel-grouping number 16.
| WAD2018 | Score | Difference |
|---|---|---|
| 1st | 86.18 | -0.09 |
| Mine | 86.09 | +0.0 |
| 2nd | 86.04 | +0.05 |
| 3rd | 84.01 | +2.08 |