DAM4SAM: Enhancing Visual Object Tracking with Distractor-Aware Memory

At CVPR 2025 paper, “A Distractor-Aware Memory for Visual Object Tracking with SAM2,” introduces a novel memory-based module that improves the state-of-the-art segmentation model SAM2.1 by effectively handling distractors — objects visually similar to the target that often lead to tracking errors.

This approach uses a smart update strategy to manage two memory buffers: one storing recent appearances for maintaining segmentation accuracy, and another holding anchor frames to resolve distractor ambiguities. This design boosts both accuracy and robustness without retraining the original SAM2 backbone.

The authors also propose DiDi, a new dataset focusing on distractor-rich scenarios, where this method achieves state-of-the-art results, outperforming SAM2.1 and other leading trackers.

Additionally, this module generalizes well, providing notable performance gains when integrated with other trackers like EfficientTAM (+11%) and EdgeTAM (+4%), demonstrating versatility and real-time capability.

Follow Me

Follow me on LinkedIn