| source | description |
|---|---|
| data set | Kitti dataset |
| FCN16.py | FCN-16 model, ResNet18 as pretrain classification model |
| FCN32.py | FCN-32 model, ResNet18 as pretrain classification model |
| KittiDataset.py | custom dataset for semantic segmentation |
| experiment.py | 1. dataset splitting 2. training and validation processing |
| test.py | test process |
| labels.py | Kitti labels definition |
Unlike original FCN setting, both FCN in this report are based on ResNet18
All image are resize to (H: 375, W: 1242) to improve parallelism.
| max_lr | epoch | weight_decay | batch size | n_class | criterion | optimizer |
|---|---|---|---|---|---|---|
| 1e-3 | 30 | 1e-4 | 10 | 34 | cross entropy | Adam |
Evaluation metrics:
- Pixel-level intersection-over-union (pIoU)
- Mean Intersection-over-Union (mIoU)
| FCN16 | FCN32 |
|---|---|
| FCN-16 | FCN-32 | |
|---|---|---|
| loss | 1.28 | 1.61 |
| mIoU | 0.46 | 0.34 |
| pIoU |