This is an official implementation of the paper "Weakly Supervised Adversarial Domain Adaptation for Semantic Segmentation in Urban Scenes" (completed in October 2017, accepted by IEEE TIP in March 2019).

We plan to finish readme.md in December 2019.

• Pytorch 0.2
  • This work is completed in October 2017. We have no plan to upgrade it to a newer Pytorch version because of limited time.
• Python2.7
• TensorboardX

• 1 Download GTA 5, SYN, Cityscapes datasets.
• 2 Generate object-level labels:
  • 2.1 Detect the foreground objects using DSOD.
  • 2.2 Get the background objects' locations according to segmentation masks.
  • 2.3 Merge the locations of fore/background objects.
  • 2.4 Generate SSD-style detection labels, a txt file. Each line contains filename number_of_objects xmin ymin xmax ymax .... Example:

  1.png 1 0 0 1 1 
  2.png 2 0 0 1 1 2 2 4 4
  ...
  

• 3 Make ROI pooling.
• 4 Train model: python run_WSAL.py.

If you use the code, please cite the following paper:

@article{gao2019scar,
  title={SCAR: Spatial-/channel-wise attention regression networks for crowd counting},
  author={Gao, Junyu and Wang, Qi and Yuan, Yuan},
  journal={Neurocomputing},
  volume={363},
  pages={1--8},
  year={2019},
  publisher={Elsevier}
}