Principal sample based learning of deep network for correlation filter tracking

Correlation filter based tracking has shown impressive and amazing results for the object tracking domain. The types of features used in this family of trackers significantly affect the performance of the tracking process. In order to achieve the significant features, deep networks can be combined with correlation trackers. In this work, the principal sample-based learning of deep networks has been proposed for correlation filter tracking. Usual training of deep networks always takes all sample frames for backpropagation, which is not ideal for large tracking video collections. To overcome these shortcomings, a novel sample selection strategy is devised while back-propagating the network loss or error, and hence the proposed method is named as Principal Sample-based Learning of Deep Network (OT-PSLDN) for correlation filter based object tracking. The proposed OT-PSLDN method is evaluated with standard performance criteria on benchmark datasets, namely Object Tracking Benchmark 2013 (OTB 2013), OTB 2015, Visual Object Tracking 2017 (VOT 2017), and VOT 2018. The experimental results show that the proposed method constantly exceeds the state-of-the-art methods in terms of qualitative and quantitative performance measures.