FCNs in the Wild: Pixel-level Adversarial and Constraint-based Adaptation Implemented by Tensorflow

Paper link: https://arxiv.org/abs/1612.02649

Intro

Tensorflow implementation of the paper for adapting semantic segmentation from the (A) Synthia dataset to Cityscapes dataset and (B) Cityscapes dataset to Our dataset.

Installation

• Use Tensorflow version-1.1.0 with Python2

• Build ccnn

   cd fcns-wild
   mkdir build
   cd build
   cmake ..
   make -j8
  

Dataset

• Download Cityscapes Dataset
• Download Synthia Dataset
  • download the subset "SYNTHIA-RAND-CITYSCAPES"
• Download NMD Dataset
  • contains four subsets --- Taipei, Tokyo, Roma, Rio --- used as target domain (only testing data has annotations)
• Change the data path in files under folder "./data"

Testing

• Download and testing the trained model

   cd fcns-wild
   sh scripts/download_demo.sh
   sh scripts/infer_city2NMD.sh 	# This shell NMD is using Taipei
  

  The demo model is cityscapes-to-Taipei, and results will be saved in the ./train_results/ folder. Also, it shows evaluated performance. (the evaluation code is provided by Cityscapes-dataset)

Training Examples

• Download the pretrained weights (model trained on source)

   sh scripts/download_src.sh
  

• Train the Cityscapes-to-Ours{subset} model

   python ./src/train_adv.py \
   	--weight_path ./pretrained/train_cscape.npy \
   	--city {city_name} \
   	--src_data_path ./data/Cityscapes.txt \
   	--tgt_data_path ./data/{city_name}.txt \
   	--method GACA \
  

  The training scripts for adapt from (A) Synthia dataset to Cityscapes dataset and (B) Cityscapes dataset to Our dataset are prepared in "scripts/run_train_syn2city.sh" and "scripts/run_train_city2ours.sh".

Reference code

https://github.com/pathak22/ccnn

https://github.com/MarvinTeichmann/tensorflow-fcn