A multi-level hybrid siamese network using box adaptive and classification approach for robust tracking

Visual tracking poses a complex challenge in computer vision, critical for applications like surveillance, autonomous vehicles, and augmented reality. In response to this challenge, our paper introduces SiamAdapt, a pioneering solution designed to redefine the landscape of visual tracking. By reshaping the traditional paradigm, SiamAdapt decomposes the tracking task into pixel-level classification and bounding box regression, providing a unique and effective approach to end-to-end visual tracking. Differentiating itself from established methods such as Siamese-RPN and SiamRPN++, SiamAdapt embraces a proposal and anchor-free methodology, eliminating the intricate process of anchor hyperparameter tuning. Utilizing a box adaptive method and inspired by the limitations of existing trackers that often resort to multi-scale searching or predefined anchor boxes, SiamAdapt excels with a siamese subnetwork for feature extraction and a classification-regression subnetwork which harnesses the power of a fully convolutional network (FCN). Additionally, we introduced two key ranking-based losses to address limitations in existing Siamese trackers. The classification ranking loss prioritizes positive samples over hard negatives, enhancing foreground detection. The IoU-guided ranking loss aligns classification with localization accuracy, boosting confidence in well-localized predictions. Through rigorous experimentation across challenging benchmarks like OTB100, UAV123, GOT-10K, VOT2018, VOT2019, and LASOT, our results demonstrate that SiamAdapt achieves leading performance, solidifying its position as a cutting-edge solution in visual tracking, particularly notable for its commendable real-time speed.