Rethinking the Grouping Process in Corner-Based Detectors
Haoran Wei, Yangguang Zhu, Ping Guo, Bing Wang, Jiamei Fu · 2022 IEEE International Conference on Multimedia and Expo (ICME) · 2022
The grouping process of corner-based detectors still faces two challenges: 1) Hard-grouping. If one of the paired corners is wrongly estimated, the grouping goes wrong. 2) Complex pipeline. Vanilla methods regard corner grouping as an additional stage, complicating the post-processing. To eliminate these issues, we propose a novel grouping algorithm, termed as Soft-Grouping Non-Maximum Suppression (SG-NMS), which merges grouping with NMS into a whole to simplify the pipeline and lift the efficiency. SG-NMS is flexible to match the varied number of detected corners. Accordingly, we propose a multi-corners context enhanced module, Corner Visual Reasoning (CVR), as a grouping helper. Equipped with the proposed SG-NMS, a new multi-corners detector, SGCDet (Soft Grouping Corner-based Detector), is proposed. Experiments show that the inference speed of SGCDet is more than two times faster than state-of-the-art corner-based models with much higher accuracy.