This repository contains the implementation for the paper:

Importance of Self-Consistency in Active Learning for Semantic Segmentation (BMVC 2020)

The code is test on Ubuntu 16.04, Pytorch 1.5, and python 3.6.6.

The Active learing code in this repository consist of two main steps (folders):

1. ActiveLearning: Include active learning code.
   
2. FinalTraining: Include the code for final training stage after AL process is finished or to compute the upperbond performance.

In addition to the resuts in the paper here is also the results via DeepLabV3 as an backbone (instead of FCN) for our paper.

Results for DeeplabV3 as backbone

Dataset/Model	Input Resolution	Classes	Batch Size	Epochs	Mean IoU (%)	Budget	Self-Consistency
Camvid (Fully Trained)-DeepLabV3	360x480	11	5	60	0.667	100%	0
Cityscapes (Fully Trained)-DeepLabV3	512x1024	19	4	60	0.649	100%	0
Camvid (Fully Trained)-DeepLabV3+	360x480	11	5	60	0.672	100%	1
Cityscapes (Fully Trained)-DeepLabV3+	512x1024	19	4	60	0.697	100%	1
Camvid (Active Learning)-DeepLabV3	360x480	11	5	60	0.622	12%	0
Cityscapes (Active Learning)-DeepLabV3	512x1024	19	4	60	0.633	12%	0
Camvid (Active Learning)-DeepLabV3+	360x480	11	5	60	0.634	12%	1
Cityscapes (Active Learning)-DeepLabV3+	512x1024	19	4	60	0.674	12%	1

Citation:

@article{equal2020,
  title={Importance of Self-Consistency in Active Learning for Semantic Segmentation},
  author={Golestaneh, S. Alireza, Kitani, Kris},
  journal={BMVC},
  year={2020}
}

If you have any questions about our work, please do not hesitate to contact us by emails at [email protected]

Acknowledgment: part of the implementation is borrowed from SegNet and Pytorch.