Predicting YOLO Misdetection by Learning Grid Cell Consensus

Bijay Raj Paudel, Danushka Senarathna, Haibo Wang, Spyros Tragoudas, Yao Hu, Shengbing Jiang

Abstract: Despite the immense performance improvement of deep learning-based object detection, the state-of-the-art object detection systems are still prone to misdetections. This work presents a method to predict such misdetections at run-time by using a small network, referred to as ConsensusNet, to learn the correlation patterns or consensus of neighboring detections before non-maximum suppression (NMS). Based on such correlations, ConsensusNet predicts if there are misdetection failures. The proposed method is experimentally evaluated considering single person class from COCO dataset and using YOLOv3 as the object detection system. It shows the proposed method can achieve accuracy of 84.6% and the performance measured in other metrics are also promising. To the best of our knowledge, ConsensusNet is the first network reported for predicting misdetections in object detection.