An Enhanced YOLO Failure Detection Method

Danushka Senarathna, Rezoan Ferdous, Spyros Tragoudas

Abstract: A novel method is proposed in this paper to detect failures in the YOLO object detection network. The proposed method is derived based on the features extracted by the YOLO network and uses a secondary neural network to predict misdetections. Subsequently, a novel Recursive Feature Elimination (RFE) based approach is proposed to make the secondary network more lightweight by selecting important features for a target class(es). Hence the computational cost is reduced with a minimum loss of accuracy. Experimental evaluation was done using a YOLO network trained on the COCO dataset considering four of the most frequently appeared classes in the dataset. The proposed failure detection method achieved an 89.79% accuracy when a single class was considered, and a 16% improvement was observed in the accuracy compared to an existing method. By using the proposed feature selection method, an 88.89% accuracy with a 56% reduction in the inference time was achieved. Feature selection was 62 times faster and achieved almost the same failure detection accuracy with a lesser number of features compared to the conventional RFE approach. Moreover, the proposed failure detection framework was evaluated by considering multiple classes together as well, and high accuracy was observed.