Object Detection for Similar Appearance Objects Based on Entropy

Abstract

In order to detect objects with similar appearance more accurately, we propose an object detection algorithm with entropy loss. Applying entropy loss makes detector predicts the class of detected bounding boxes more robust with high score probability. It also leads to decrease of confidence loss. Therefore, the detection performance for similar objects is improved. We reconstructed the dataset from previous two datasets to evaluate our method, implemented experiments, and obtained high performance gain. In addition, we conducted an analysis of the score distribution for detected objects and the other loss terms, in order to observe the effects of applying entropy loss.