Reinforced active learning for image segmentation

Abstract

Learning-based approaches for semantic segmentation have two inherent challenges. First, acquiring pixel-wise labels is expensive and time-consuming. Second, realistic segmentation datasets are highly unbalanced: some categories are much more abundant than others, biasing the performance to the most represented ones. In this paper, we are interested in focusing human labelling effort on a small subset of a larger pool of data, minimizing this effort while maximizing performance of a segmentation model on a hold-out set. We present a new active learning strategy for semantic segmentation based on deep reinforcement learning (RL). An agent learns a policy to select a subset of small informative image regions -opposed to entire images -to be labeled, from a pool of unlabeled data. The region selection decision is made based on predictions and uncertainties of the segmentation model being trained. Our method proposes a new modification of the deep Q-network (DQN) formulation for active learning, adapting it to the large-scale nature of semantic segmentation problems. We test the proof of concept in CamVid and provide results in the large-scale dataset Cityscapes. On Cityscapes, our deep RL region-based DQN approach requires roughly 30% less additional labeled data than our most competitive baseline to reach the same performance. Moreover, we find that our method asks for more labels of under-represented categories compared to the baselines, improving their performance and helping to mitigate class imbalance.

Active learning for object detection in high-resolution satellite images

“…Interestingly, some recent approaches have tried to treat the active learning problem as a reinforcement learning task. In [4], the authors propose to train a reinforcement learning algorithm to select the most useful patches in images from two public datasets -CamVid and Cityscape -to improve an existing segmentation model. Their solution is not straightforward to implement and remains more computation heavy than the others presented here but their results on both datasets seem to outperform other uncertainty-based approaches.…”

Embodied Learning for Lifelong Visual Perception

“…The approach by Kasarla et al (2019) uses entropy-and region-based AL, with annotation at the superpixel level in images, along with the use of fully connected CRFs for label propagation. Mackowiak et al (2018) study region-based AL by estimating labelling cost and uncertainty of unlabelled regions, and Casanova et al (2020) present a deep reinforcement learning-based approach, where an agent learns a policy to select a subset of image patches to label.…”

A Comparative Survey of Deep Active Learning

“…detection [Haussmann et al, 2020;Roy et al, 2018], image segmentation [Casanova et al, 2020;Saidu and Csató, 2021], counting [Zhao et al, 2020], etc. Besides these applications, some researchers have designed unified DAL frameworks that perform well on various tasks like [Ash et al, 2019;Pinsler et al, 2019;Sener and Savarese, 2017;Shui et al, 2020].…”