Accurate bounding-box regression with distance-IoU loss for visual tracking

Most existing trackers are based on using a classifier and multi-scale estimation to estimate the target state. Consequently, as expected, have become more stable while tracking accuracy has stagnated. While adopt maximum overlap method an intersection-over-union (IoU) loss mitigate this problem, there defects in IoU itself, that make it impossible continue optimize objective function when given bounding box is completely contained within/without another box; makes very challenging accurately Accordingly, paper, we address above-mentioned problem by proposing novel distance-IoU (DIoU) loss, such proposed tracker consists of classification. The part trained predict DIoU score between ground-truth bounding-box estimated bounding-box. can maintain advantage provided minimizing distance center points two boxes, thereby making accurate. Moreover, introduce classification online optimized with Conjugate-Gradient-based strategy guarantee real-time speed. Comprehensive experimental results demonstrate achieves competitive compared state-of-the-art