Siamese target estimation network with AIoU loss for real-time visual tracking.

The fully convolutional siamese network based trackers achieve great progress recently. Most of these methods focus on improving the capability of siamese network to represent the target. In this paper, we propose our model which focuses on estimating the state of the target with our proposed novel IoU (intersection over union) loss function which is named AIoU. Our model consists of a siamese subnetwork for feature extraction and a target estimation subnetwork for state representation. The target estimation subnetwork contains a classification head for classifying background and foreground and a regression head for estimating target. In order to regress better bounding boxes, we further study the loss function utilized in the regression head and propose a powerful IoU loss function. Our tracker achieves competitive performance on OTB2015, VOT2018, and VOT2019 benchmarks with a speed of 180 FPS, which proves the effectiveness of our method. • Powerful state estimation approaches promote Siamese trackers' performance. • Visual tracking can be decomposed to classification and regression task. • IoU loss with appropriate penalty terms leads to better regression results. [ABSTRACT FROM AUTHOR]