Skip to content

jonasdieker/toy-segmentation

Repository files navigation

Segmentation Benchmark

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.

Results

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

Getting Started

  • 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

About

Segmentation experiments: U-Net, DeepLab, Segformer

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published