This is an implementation of ENet: A Deep Neural Network Architecture for Real-Time Semantic Segmentation, ported from ENet-training (lua-torch) to keras.

git clone https://github.com/PavlosMelissinos/enet-keras.git
cd enet-keras

On poetry: poetry install

On Anaconda/miniconda: conda env create -f environment.yml

On pip: pip install -r requirements.txt

make setup

The setup script only sets up some directories and converts the model to an appropriate format.

make train

• Clean up code
  • Remove hardcoded paths
  • Add documentation everywhere
• Test code
  • Add tests
• Fix performance (mostly preprocessing bottleneck)
  • Remove unnecessary computations in data preprocessing
  • Index dataset category internals. Dataset categories have fields with one-to-one correspondence like id, category_id, palette, categories. This seems like perfect table structure. Might be too much though.
  • (Optionally) Make data loader multithreaded (no idea how to approach this one, multithreadedness is handled by keras though)
• Enhance reproducibility/usability
  • Upload pretrained model
  • Finalize predict.py
    • Test whether it works after latest changes
    • Modify predict.py to load a single image or from a file. There's no point in loading images from the validation set.
• Fix bugs
  • Investigate reason for bad results, see #11
  • Fix MSCOCOReduced, also see #9
  • ?????