Scholarlink

<p>RGB-only visual object tracking (VOT) degrades under fast motion, motion blur, and extreme illumination, while event cameras offer microsecond latency and high dynamic range but lack rich spatial semantics. We introduce UREPTrack, a unified, single-stage, attention-free RGB-event tracker built on a lightweight PoolFormer backbone. Raw events are voxelized into compact spatiotemporal tensors and, together with RGB template/search patches, embedded and concatenated into a single token stream processed by a shared backbone; a fully convolutional head jointly predicts classification confidence, center offsets, and box size. This design removes multi-branch Siamese pipelines and costly self-attention, delivering strong accuracy at very high speed. UREPTrack sets a new state-of-the-art (SOTA) on COESOT with S 64.4, P 77.5, NP 76.2, and BOC 23.7 while running at 170 FPS. On VisEvent, it achieves S 55.46, SR<sub>0.5</sub> 67.01, SR<sub>0.75</sub> 46.96, P 71.58, and NP 75.22, and on FE108 it reaches P 94.3 and S 65.9. Ablations confirm (i) the benefit of RGB and event complementarity, (ii) the superiority of event voxels over image-like alternatives, and (iii) favorable accuracy, efficiency scaling across PoolFormer sizes. UREPTrack offers a practical path to real-time, multi-modal tracking on resource-constrained hardware.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

UREPtrack: single-branch PoolFormer for unified attention-free RGB-event visual object tracking

  • 期刊论文

  • 发表日期：

    2025年12月17日

查看全文
  • Min Lu

摘要

RGB-only visual object tracking (VOT) degrades under fast motion, motion blur, and extreme illumination, while event cameras offer microsecond latency and high dynamic range but lack rich spatial semantics. We introduce UREPTrack, a unified, single-stage, attention-free RGB-event tracker built on a lightweight PoolFormer backbone. Raw events are voxelized into compact spatiotemporal tensors and, together with RGB template/search patches, embedded and concatenated into a single token stream processed by a shared backbone; a fully convolutional head jointly predicts classification confidence, center offsets, and box size. This design removes multi-branch Siamese pipelines and costly self-attention, delivering strong accuracy at very high speed. UREPTrack sets a new state-of-the-art (SOTA) on COESOT with S 64.4, P 77.5, NP 76.2, and BOC 23.7 while running at 170 FPS. On VisEvent, it achieves S 55.46, SR0.5 67.01, SR0.75 46.96, P 71.58, and NP 75.22, and on FE108 it reaches P 94.3 and S 65.9. Ablations confirm (i) the benefit of RGB and event complementarity, (ii) the superiority of event voxels over image-like alternatives, and (iii) favorable accuracy, efficiency scaling across PoolFormer sizes. UREPTrack offers a practical path to real-time, multi-modal tracking on resource-constrained hardware.