SAM-STIR: Detection-Guided Memory Update for SAM 3-Based Infrared Small and Tiny Object Tracking

Infrared small and tiny object tracking plays an important role in security surveillance, remote sensing, and military applications. The vision foundation-model segment anything model (SAM) 3 demonstrates strong tracking capability through a cooperative verification mechanism between its detector and tracker, yet its performance degrades substantially on infrared small and tiny targets. We identify tracker-drift-induced memory collapse as a central failure mode: for small and tiny infrared targets, drifted self-propagated masks can be repeatedly written into the memory bank, gradually turning a reliable temporal prior into a source of self-reinforcing errors. To counter this memory collapse, we propose SAM-STIR, a SAM 3-based infrared tracking framework centered on detection-guided memory update. Instead of relying on the tracker itself as the memory-update source, SAM-STIR refreshes memory with matched high-confidence detector masks, thereby reducing drift-induced memory contamination. To make this update reliable for small and tiny targets, we introduce a normalized distance-based matching mechanism as a geometric matching gate before memory reconditioning. We further employ a lightweight spatio-temporal differential aggregation (STDA) front-end to provide infrared contour and motion cues for subsequent matching and propagation. Experiments on the CST antiuncrewed aerial vehicles (UAV) and Anti-UAV 410 datasets demonstrate that the proposed method achieves the best mSA among the compared methods, reaching 62.63% and 69.41%, respectively, and outperforming the fine-tuned SAM 3 baseline by 9.43 and 4.59 percentage points.