ResamplingNet: End-to-End Adaptive Feature Resampling Network

Object feature pollution is one of the burning issues in UAV tracking, which is commonly caused by occlusion, fast motion, and illumination variation. Due to the contaminated information in the polluted object features, most trackers fail to precisely estimate the location and scale of the object. To address the feature pollution issue, this work proposes an efficient and effective adaptive feature resampling tracker, \textit{i.e.}, AFRT. AFRT mainly includes two stages: an adaptive downsampling network which can reduce the interference information of the feature pollution and a super-resolution upsampling network, applying Transformer to restore the object scale information. Specifically, the adaptive downsampling network strengthens the expression of the object location information, with a feature enhancement downsampling (FED) module. In order to achieve better training effect, a novel pooling distance loss function is designed to help FED module focus on the critical regions with the object information. Thereby, the features downsampled can be validly exploited to determine the location of the object. Subsequently, the super-resolution upsampling network raises the scale information in the features with a low-to-high (LTH) Transformer encoder.