This is the code for the following paper:
Cai, Lile, Xun Xu, Jun Hao Liew, and Chuan Sheng Foo. "Revisiting Superpixels for Active Learning in Semantic Segmentation With Realistic Annotation Costs." In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10988-10997. 2021.
The code is tested with Tensorflow-1.13.2 with Python 3.6.8 using docker image tensorflow/tensorflow:1.13.2-gpu-py3.
-
Prepare Cityscapes dataset. The Cityscapes dataset should be put in ./deeplab/datasets/cityscapes. You may refer to the deeplab repo (https://github.com/tensorflow/models/tree/master/research/deeplab) for the details.
-
Prepare the xception-65 model pretrained on ImageNet. The pretrained model should be put in ./deeplab/models. You can download the weights from the deeplab model zoo.
-
Run python ./scripts/extract_superpixels.py to extract superpixels.
-
Run python ./scripts/extract_rectangles.py to extract rectangles.
-
Run python ./scripts/compute_anno_cost.py to compute the annotation cost for each rectangle using polygon-based annotation.
-
Run ./scripts/write_bash_files.py to generate the bash files. Then run: bash ./bash_files/job_name.sh to run the experiment.
'Sp+Do+Random': region_type = 'sp', v0
'Sp+Do+Uncertainty': region_type = 'sp', v1, is_bal = False
'Sp+Do+ClassBal': region_type = 'sp', v1, is_bal = True
'Rec+Pr+Random': region_type = 'rec', v0
'Rec+Pr+Uncertainty': region_type = 'rec', v1, is_bal = False
Here is the output for 1 run:
