Object Tracking Via Adaptive Frequency Domain Filter and Condensation Loss
SUN Peisheng
FAN Jiaqing
SONG Huihui
Abstract：Visual object tracking(VOT)is a fundamental task in computer vision,which has widely applied in realistic sce-narios.Therefore,the research on visual object tracking is of great significance.There are three unsolved problems in previous VOT algorithms.Firstly,with the proposal of ResNet,spatial feature extraction is greatly strengthened,but there inherently have several drawbacks.Secondly,for the extracted features,how to leverage useful information and suppress useless information is still chal-lenging for VOT.Thirdly,the sample imbalance leads to the problem that the model lacks the discriminative ability,which limits the performance of the learned model.In order to solve the issues above,this paper uses the dual feature enhancement module to en-hance the spatial features extracted by ResNet.The spatial domain features are transformed into frequency domain features by the Fourier transform,and then the suitable features for the tracker are automatically highlighted by employing the learnable filter in the frequency domain,while suppressing other chaotic background information.Finally,the aggregation loss is introduced to expand the range of difficult samples and punish simple samples,which significantly alleviates the problem of sample imbalance.A large number of experimental results on four challenging datasets show that the proposed method has favorable performance against SOTA algorithms.In particular,the proposed approach achieves a AUC score of 62.0%in the TLP and 1.7%improvement in the VOT2020LT.
Keywords：object trackingdual feature enhancement moduleadaptive frequency domain filtersample imbalancecoagu-lation loss
Publication Date：2025-03-20
Online Publishing Date：2025-08-15(First online date of this platform, not the publication date of the document)
Pages：9( 725-733 )