Visual Object Tracking Based On Meta Learning

Existing deep trackers can be divided into two categories: trackers without online update and those with online update.The first category relies on massive training data to obtain a robust network.These methods need no online updating,thus,they usually run at real-time speeds.While,these methods cannot adapt to appearance variations of target without important online adaptability,thereby increasing the risk of tracking drift.Algorithms in the second group improve the discriminative ability of deep networks by frequent online update.They utilize the first frame to initialize the model and update it every few frames.Timely online update enables trackers to capture target variations but also requires more computation.Therefore,the speed of these trackers generally cannot meet the real-time requirements.To take full of the advantages of the two algorithms and avoid their disadvantages,this thesis proposes the Dynamic Matching-Classification Switching(MCS)framework which integrates the matching and classification networks.A verification network is also used to conduct dynamic switching between them.In simple tracking situation,the matching-based network is used to track.When there is a lot of noisy,the classification-based network starts to work.The dynamic switching between them leads the limited computation to the complex situations.To fast the online update of the classification network,a meta-classifier is further proposed to realize the fast adaptation through one iteration.Extensive experiments on general benchmarks demonstrate the effectiveness of the proposed method.Although the tracking method based on Dynamic Matching-Classification Switching achieves good performance,the model size is very huge,which makes it hard to be applied in the real world.To reduce the parameters of MCS,we further propose a Gradient-Guided Network(GradNet)which exploits the discriminative information in gradients and updates the template in Siamese network through feed-forward and backward operations.Besides,a template generalization training method is proposed to force the updater to focus on the gradients and avoid overfitting.Experiments on popular benchmarks demonstrate that our algorithms achieve much better performance than other state-of-the-art trackers.