A github version of kaggle notebook where I explored few semantic segmentation architectures on subset of CityScapes dataset. The notebook contains IoU, DICE losses, and 4 architectures :
- UNet,
- ENet,
- ESNet,
- DeepLabV3.
Link to kaggle notebook https://www.kaggle.com/code/levusik/semantic-segmentation-pytorch-unet-enet-esnet
Summary of results
| Network Name | mDICE | mIoU | number of parameters |
|---|---|---|---|
| UNet | 0.8541 | 0.7463 | 7,763,272 |
| ENet | 0.835 | 0.716 | 350,076 |
| ESNet | 0.8515 | 0.7422 | 1,658,356 |
| DeepLabV3 | 0.8617 | 0.758 | 39,635,528 |
