Siamese visual tracking based on criss-cross attention and improved head network

The efficient Siamese anchor-free tracker has fewer parameters, but it produces a large number of low-quality bounding boxes which are located far away from the center of the object. Moreover, a plenty of background information or distractors also interfere with the tracking process, resulting in the inaccurate results of classification and regression. As such, we propose a novel Siamese anchor-free network based on criss-cross attention and an improved head network. We apply ResNet-50 to extract the features of the template image and search region, then feed the feature maps into a recurrent criss-cross attention module to make it more discriminative. The enhanced feature maps are inputted into our improved head network, which include the center-ness branch based on the original classification and regression branches to filter out low-quality bounding boxes. Our proposed tracker reduces the impact of background information or distractors and can obtain high-quality bounding boxes, generating more accurate and robust tracking results. Extensive experiments and comparisons with state-of-the-art trackers are conducted on many challenging benchmarks such as VOT2016, VOT2018, GOT-10k, UAV123 and OTB2015. Our tracker achieves excellent performance with a considerable real-time speed.