Improving Real-world Object Detection Using Balanced Loss

Abstract

Training process is crucial to the success of object detectors. Real-world datasets often have skewed distributions, which results in imbalance issues during the training process, and thus affects the performance of detector. In this work, we propose a simple but effective framework towards balanced learning for object detection, in which two balanced loss functions are developed to alleviate the imbalance during training: balanced cross entropy loss and balanced classification regression loss, respectively for reducing the imbalance at positive-negative samples and objective level. Experiments on one pedestrian detection and one foreign particle inspection task show that, benefitted from the balanced loss design, the detection performance can be improved significantly.