This repo serves as an experiments to compare CNN-based segmentation networks to the transformer-based SegFormer model.
A goal was also to make getting started with SegFormer eaiser, by not using other repo dependencies such as mmcv or the like.
Without doing too much hyper-parameter tuning these are the results obtained on the CamVid dataset:
| Model | Dice Score | mIoU |
|---|---|---|
| U-Net | 0.72 | 0.34 |
| DeepLabV3 | - | - |
| SegFormer (B0) | 0.77 | 0.27 |
-
Download the data from here
-
Create environment and install requirements:
conda create -n segmentation python=3.9 conda activate segmentation pip install -r requirements.txt -
Run training script
train.py