On uncertainty estimation in active learning for image segmentation

This repository provides the implementation for our paper On uncertainty estimation in active learning for image segmentation (Bo Li, Tommy Sonne Alstrøm). We experimentally show that the region based active learning strategy can lead to higher segmentation accuracy and better calibrated model much faster than full image acquisition:

Installation and preparation

1. Clone and enter this repo:

   git clone https://lab.compute.dtu.dk/papers/on-uncertainty-estimation-in-active-learning.git
   cd on-uncertainty-estimation-in-active-learning
   chmod +x requirement.sh
   chmod +x produce_figure.sh

2. Create a virtual env with the required packages

   conda env create -f active_learning.yaml
   source activate act

3. Prepare the Dataset and pertained resnet-50 ckpt

   ./requirement.sh

Evaluate the model

In order to evaluate the model at each acquisition step, run

python3 -c 'import Test as te;te.running_test_for_single_acquisition_step(model_dir)'
Args:
  model_dir: the directory that saves the model ckpt

Train the model

• For full image based active learning, run

  python3 -c 'import Train_Active_Full_Im as tafi;tafi.running_loop_active_learning_full_image(stage)'
  Args:
    stage: int, 0:random, 1:VarRatio, 2:Entropy, 3:BALD

• For region based active learning, run

  python3 -c 'import Train_Active_Region_Im as tari;tari.running_loop_active_learning_region(stage)'
  Args:
    stage: int, 0:random, 1:VarRatio, 2:Entropy, 3:BALD

Reproduce figures

The statistics that are used for reproducing the figures are saved in folder Exp_Stat. In order to reproduce the figures in the paper, run

./produce_figure.sh