Domain Adaptive Visual Tracking with Multi-scale Feature Fusion

Accuracy and speed have been two fundamental issues that are difficult to balance in object tracking. Trackers with high accuracy often have quite large network structures that require huge amounts of computing resources, therefore leading to a lower tracking speed. To address the problem, we propose a novel domain adaptive tracking algorithm to obtain a better balance between tracking speed and accuracy. A simple and effective domain adaptation component is employed to transfer features from the image classification domain to the object tracking domain. In addition, we construct an adaptive spatial pyramid pooling layer to substitute for the fully- connected layer connected to convolutional layers, which can significantly reduce computational complexity while achieving high tracking accuracy. Experiments on VOT2018, TrackingNet and OTB2015 shown the effectiveness of the proposed method. Compared with the state-of-the-art trackers, our tracker can obtain real-time tracking with a speed of 35 FPS.