Crowd Counting From Limited Labeled Data Using Active Learning

Deep learning models have provided dramatic performance improvement for various computer vision tasks. These models, however, require huge amounts of labeled data to perform well. Collecting and labeling large datasets is often non-trivial and requires significant human effort. Crowd counting is one such task that demands a large amount of labeled training data. This labeling process requires a human annotator to manually mark a dot at the center of the head of each person present in the image, which is a laborious and tedious task, especially in densely crowded scenes. In this work, we investigate an active learning framework for crowd counting. Evaluations on mainstream datasets demonstrate the effectiveness of the proposed framework in reducing the annotation effort significantly with minimal compromise on count performance. Our method surpasses existing methods that focus on counting with limited labeled data.