This is the authors' abstract. We don't add key points for this paper.
Infrared small target detection remains challenging because the targets provide weak appearance cues and are easily submerged by background clutter, sensor noise, and interference induced by platform motion. Recent studies have advanced the field from both single-frame and multiframe perspectives, showing that temporal information can complement insufficient single-frame evidence. However, the practical gain of multiframe processing is still constrained by the stability of interframe correspondence. When motion estimation is unreliable, temporal accumulation amplifies misalignment-induced residuals together with target responses. To better exploit temporal information under imperfect interframe correspondence, we propose the multiframe infrared spatiotemporal aggregation (MISA) framework, which combines region-based temporal prior generation (RTPG) with asymmetric fusion. Specifically, MISA introduces a RTPG module, which replaces isolated point support with multiregion joint tracking to generate a more reliable temporal prior from multiple support regions. To mitigate interference between temporal priors and appearance features, MISA further introduces a Prior-guided Asymmetric Fusion (PAF) module that uses asymmetric interactions to preserve temporal cues during detection. Experiments on the IRDST and third anti-UAV datasets show that MISA achieves competitive performance among recent online methods while maintaining high efficiency. Specifically, MISA-R18 attains 95.1% mAP ₅₀ on IRDST with 57.5G FLOPs, while the lightweight MISA-YOLO11 variant reaches 93.0% mAP ₅₀ with 25.5 G FLOPs.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Wu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: