Split-merge-excitation: a robust channel-wise feature attention mechanism applied to MDNet tracking

Han Wu,Guizhong Liu
DOI: https://doi.org/10.1007/s11042-022-12752-z
IF: 2.577
2022-01-01
Multimedia Tools and Applications
Abstract:Object tracking is a fundamental problem of computer vision. Although being studied for decades, the single object tracking problem has not been completely solved, since there exist various challenges in the real physical world, such as object deformation, complex background and imperfect imaging, which make tracking difficult. For these challenges, we design a robust feature extraction network. Specifically, we propose a novel channel-wise feature attention mechanism, which is integrated into the pipeline of a well-known convolutional neural network based visual tracking algorithm. It is crucial to represent the object robustly. Due to the representative feature, the tracking performance is improved. In experiments, we test the proposed tracking algorithm in OTB100, VOT2018, VOT2020 and VOT-TIR datasets. Compared to the baseline algorithm, our proposed algorithm obtains consistent performance improvement for different benchmarks with absolute increase of tracking success score in OTB100 up to 0.6, and absolute increase of EAO up to 0.022, 0.007, and 0.008 in VOT2018, VOT2020, VOT-TIR2015 respectively. The source codes are publicly available.
What problem does this paper attempt to address?