Published: Jan. 1, 2023

Deep learning-based object tracking networks have gained prominence due to their accuracy and speed, currently use two main frameworks: dual-stream two-stage single-stream single-stage tracking. The former ignores information interaction during feature extraction, the latter overlooks backbone network's own limitations. To address these issues, we adopted framework proposed optimizations for network. First, propose an objective contour enhancement module distinguish similar objects from background through high-pass filtering. Secondly, introduce a patch fusion module, which employs learnable tensors within frequency domain establish relationships among isolated edge patches, enabling between non-overlapping patches. Furthermore, multi-scale based on grouped shuffling mechanism, endows network with ability learn while only introducing lower parameter computational burden. Lastly, utilize DropMAE pre-trained model further enhance generalization ability. Our tracker achieves relatively outstanding performance multiple benchmark datasets. Specifically, AO metric of GOT-10k dataset, it outperforms TATrack-B SeqTrack-B384 by 3.4% 1.9%, respectively. is released https://github.com/Nirvanalll/EFTrack.

Language: Английский