Size-Sensitive Optimization of Loss Function on Vision-based Object Detection

Vision-based object detection is constantly seeking a solution of higher accuracy and faster speed. Loss function of object detection algorithm directly influences a lot on backpropagation of training and counts to the result of detection. To design a better loss function, different sizes of bounding boxes are much more considered in this paper. In detail, instead of the absolute distance, the location loss with the relative distance of bounding box sizes is calculated and an L2 regularization is added to the loss function. It is trained with VOC datasets and finally it comes out to the 58 frames per second (FPS) detection speed with 416×416 image and 82.38% mean average precision (mAP) on VOC validation dataset with the network backbone of Darknet-53. It is proved efficiently in surrounding information lost and overlapped daily objects detection and has good performance in real-time detection with 720p webcam.