Deep Siamese Network with Co-channel and Cr-Spatial Attention for Object Tracking

Siamese trackers with offline training strategies have recently drawn great attention because of their balanced accuracy and speed. However, some limitations still remain to overcome, i.e., trackers cannot robustly discriminate target from similar background so far. In this paper, we propose a novel real-time co-channel and spatial attention based deeper Siamese network (DCANet). Our approach aims at dealing with some challenging situations like appearance variations, similar distractors, etc. Different from replacing the backbone network Alexnet with VGG16 directly, we modified the structure of VGG16 which has no fully connective layer and padding operation. In addition, co-channel and spatial attention mechanisms were applied to our method to enhance feature representation capability. Channel attention and spatial attention were proposed towards computer vision problems before. However, considered the special structure of siamese network, we designed Co-channel attention module which helps to emphasize the important areas in the two branches simultaneously. When we directly add spatial attention to our tracker, the tracking effect falls. However with a crop operation placed after spatial attention our tracker can tracking better. We perform extensive experiments on several benchmark datasets, including OTB-2013, OTB-2015, VOT-2017, LaSOT and GOT-10k, which demonstrate that our DCANet gains a competitive tracking performance, with a running speed of more than 60 frames per second.