Detection Accuracy Improvement on One-Stage Object Detection Using Ap-Loss-Based Ranking Module and Resnet-152 Backbone

Abstract

Localization-loss and classification-loss are optimized at the same time to train the one-stage object detectors. Because of the large number of anchors, the severe foreground–background class disproportion causes significant classification-loss. This paper discusses using a ranking module instead of the classification module to mitigate this difficulty and also Average-Precision loss (AP-loss) is utilized on the ranking module. An optimization algorithm is used to make the AP-loss as effective as possible. Optimization algorithm blends the error-driven updating method of perceptron learning and the deep network backpropagation technique. This optimization algorithm handles the foreground–background class disproportion issues. One-stage detector with AP-loss and backbone with ResNet-152 attains improvement in the detection performance compared to the classification-losses-based detectors.