Learning non-maximum suppression
This addresses a critical bottleneck in object detection pipelines for computer vision applications, offering a learned alternative to a traditionally hand-crafted component, though it appears incremental as it builds on existing end-to-end learning paradigms.
The paper tackles the problem of non-maximum suppression (NMS) in object detection, which is a hand-crafted post-processing step that forces a trade-off between recall and precision, by proposing a new neural network architecture to perform NMS using only boxes and their scores, resulting in improved localization and occlusion handling on datasets like PETS and COCO.
Object detectors have hugely profited from moving towards an end-to-end learning paradigm: proposals, features, and the classifier becoming one neural network improved results two-fold on general object detection. One indispensable component is non-maximum suppression (NMS), a post-processing algorithm responsible for merging all detections that belong to the same object. The de facto standard NMS algorithm is still fully hand-crafted, suspiciously simple, and -- being based on greedy clustering with a fixed distance threshold -- forces a trade-off between recall and precision. We propose a new network architecture designed to perform NMS, using only boxes and their score. We report experiments for person detection on PETS and for general object categories on the COCO dataset. Our approach shows promise providing improved localization and occlusion handling.