Balanced Sampling-Based Active Learning for Object Detection

Abstract

Active learning focuses on building competitive models using less data by utilizing intelligent sampling, thereby reducing the effort and cost associated with manual data annotation. In this paper an active learning method based on a balanced sampling of images with high and low confidence object detection scores for training an object detector is proposed. Images with higher object prediction scores are sampled using the uncertainty measure proposed by Yu et al. which utilizes detection, classification and distribution statistics. Though this method encourages balanced distribution in sampling, a deeper look into the sampled distribution reveals that the under-represented classes in the initial labeled pool remain skewed throughout the subsequent active learning cycles. To mitigate this problem, in each active learning cycle, we propose to sample an equal proportion of images with high and low confidence object prediction scores from the model trained in the last cycle, where the low confidence prediction sample selection is based on the model’s prediction scores. Experiments conducted on the UEC Food 100 dataset show that the proposed method performs better than the baseline random sampling, CALD and low confidence prediction sampling method by +4.7, +8.7, and +3.1 mean average precision (mAP), respectively. Moreover, consistently superior performance of the proposed method is also demonstrated on the PASCAL VOC’07 and PASCAL VOC’12 datasets.

Work done as intern at Swiggy.