Deeper Siamese network with multi-level feature fusion for real-time visual tracking

In recent years, using Siamese network (SiamN) for visual tracking has witnessed a great success in terms of accuracy and efficiency. Nevertheless, most SiamN-based trackers employ shallow network such as AlexNet to extract the top-layer features as target representation that are less discriminative, usually leading to tracking performance degeneration when suffering from large deformation and similar distractors. A straightforward idea to address this issue is to replace the backbone network of SiamN with deeper ResNet. However, this cannot boost performance much due to the low resolution of high-level feature maps with useful spatial details losing. To address this issue, the authors propose a lightweight yet effective feature agglomeration module (FAM) to adaptively fuse low-level and high-level features for robust tracking. Specifically, they first develop a generalised non-local attention module to enhance the discriminative capability of high-level semantic features. Then, they design an inception-like module to enhance the representative power of low-level features with more spatial details. Both types of features are then adaptively fused in the FAM to complement their characteristics. Extensive evaluations on OTB-2015 and VOT2017 challenge demonstrate that the proposed tracker consistently achieves favourable performance against several state-of-the-art trackers and runs at 50 fps.