Active Learning for Semantic Segmentation with Area Disagreement

We propose Area Disagreement, a novel uncertainty estimation method for Active Learning in the context of Semantic Segmentation, which capitalizes on the unique characteristics of the task. It relies on the assumption that slight alterations to the learner model's parameters should not produce significant differences in the output, specifically in terms of predicted shapes and objects given the same input image. While a small amount of contradiction is natural, larger inconsistencies suggest that the model's internal representation of the world is perplexed by those images, thus making them valuable for further training. Our uncertainty estimation method prioritizes images with a smaller Dice Coefficient relative to the average prediction, based on multiple Monte-Carlo Dropout inferences. Utilizing this approach, we outperformed baseline methods by a wide margin on the Cityscapes dataset, achieving 95% of the full-scale training performance using only 36% of the dataset and 97.5% of the full-scale training performance using 47 % of the dataset.