SpinTrack: Cyclic token permutation for transformer-based tracking

doi:10.1016/j.neucom.2026.134612

SpinTrack: Cyclic token permutation for transformer-based tracking

2026-08-01

0

PRE

AI

DOI：10.1016/j.neucom.2026.134612

Original

Original request for help

Save

Abstract

En 中文

Transformer-based visual trackers have revolutionized the field by modeling robust long-range dependencies. However, they face a fundamental trade-off: global attention mechanisms suffer from quadratic computational complexity, while efficient window-based alternatives often restrict the receptive field, lacking the specific induc tive bias required to localize dynamic targets undergoing large spatial displacements. To address this limitation, we introduce Spin Window Attention (SWA). Unlike conventional vision Transformers that primarily use window shifting for image continuity, SWA is specifically designed for tracking. It efficiently fuses template and search region features through a strategic cyclic rotation of query tokens. By employing this rotation, the model per forms a spatial scan. The scanning enables the tracker to capture targets that move across windows, which in turn maintains linear computational complexity. Building on SWA, we propose SpinTrack, a Transformer-based tracker that achieves an optimal balance between accuracy and speed. To facilitate token rotation, an adaptive positional encoding scheme is developed to handle spatial permutations and varying resolutions-a design that further improves robustness. Extensive experiments show that SpinTrack achieves an optimal accuracy-speed trade-off on multiple benchmarks, attaining a highly competitive average overlap (AO) of 75.9% on GOT-10k while running at 75 FPS on a single RTX 4090 GPU.

Keywords:

Visual object tracking

Siamese network

Attention

Spin window attention

Transformer

SpinTrack: Cyclic token permutation for transformer-based tracking

Abstract

Journal

Organization

Cited Papers