计算机视觉
人工智能
计算机科学
可视化
图像处理
眼动
图像分割
跟踪(教育)
数据压缩
噪音(视频)
目标检测
特征提取
视频跟踪
模式识别(心理学)
稳健性(进化)
信号处理
机器视觉
跟踪系统
特征(语言学)
杂乱
人类视觉系统模型
直方图
作者
Xiaohuan Lu,Shenrun Ding,Yuji Wang,Yue Li,Xiaohuan Lu,Youliang Tian
标识
DOI:10.1109/tcsvt.2026.3690345
摘要
Visual tracking faces challenges in effectively leveraging long-term video-level contextual information due to error accumulation caused by appearance changes. Recent SAM2-based trackers address this through improved memory management but share a critical limitation: they lack mechanisms to actively detect and rectify tracking errors, allowing localization errors to gradually contaminate the memory. To address this, we propose AS-SAM2, an adaptive self-correction framework that blocks error propagation at three complementary levels by exploiting the capacity of SAM2 to accept various prompts at any frame. At the tracking level, we propose an Appearance Evaluation-based Self-correction (AES) module to identify frames with tracking errors and adopt prompts to correct them, ensuring accurate features in the memory bank for subsequent frames. At the prompt level, to mitigate background interference from box prompts, we propose a Dual-Prompt Denoising (DPD) module that employs negative prompts to enhance foreground-background distinction, thereby improving tracking robustness. At the memory level, to prevent unreliable frames from degrading the memory bank during occlusion, we propose a State-aware Adaptive Memory Selection (SAS) that dynamically switches between short-term and long-term memory based on visibility state estimation. Experimental evaluations on six benchmark datasets demonstrate that AS-SAM2 achieves state-of-the-art results on LaSOT, LaSOText, GOT-10k, and TNL2K without requiring retraining or fine-tuning, outperforming fully supervised trackers in complex scenarios. This zero-shot capability underscores its potential for deployment in real-world dynamic environments where adaptability and robustness are critical. All the source code will be released at https://github.com/oatlvv/AS-SAM2.
科研通智能强力驱动
Strongly Powered by AbleSci AI