This is the code for the paper:

Convolutional Scale Invariance for Semantic Segmentation - GCPR 2016 (Oral)

We provide:

• Training code
• Model definitions
• Implementations of weighted cross-entropy loss funcion and scale-selecion layer
• Dataset preparation code
• Code for multi-threaded data prefetching useful if the whole dataset can't fit into RAM
• Notebooks for evaluation and visualization

How to reproduce:

1. Download KITTI or Cityscapes dataset
2. Prepare the data for scale-invariant or single-scale model
3. Modify all data paths for your system and run the training script:

th train.lua -model models/scale_invariant.lua -solver solver\_config.lua

If you find this code useful in your research, please cite:

@inproceedings{kreso16gcpr,
  title={Convolutional Scale Invariance for Semantic Segmentation},
  author={Krešo, Ivan and Čaušević, Denis and Krapac, Josip and Šegvić, Siniša},
  booktitle={German Conference on Pattern Recognition},
  year={2016},
  organization={Springer}
}