Transferability of the forgetting mechanism and SSH replay

Determine whether the scale-conditioned forgetting mechanism and Scale-Stratified Herding replay transfer from YOLOMG to other YOLO variants and transformer-based object detectors.

Background

The empirical study evaluates only one architecture: YOLOMG, a YOLOv5-based detector operated with a single thermal-infrared input stream. Consequently, the observed relationship between scale distribution, gradient imbalance, and catastrophic forgetting may depend on YOLOMG’s architecture, anchors, detection heads, or normalization behavior.

The proposed cosine diagnostic is described as architecture-agnostic, but the authors explicitly leave unresolved whether the underlying finding and the SSH mitigation generalize to different detector families, including other YOLO models and transformer detectors.

References

Finally, scalability was tested on a single architecture (YOLOMG); whether the finding and SSH transfer to other YOLO variants or transformer detectors is unknown, though the cosine diagnostic itself is architecture-agnostic.

— Catastrophic Forgetting in Sequential Thermal Anti-UAV Detection: The Role of Scale-Conditioned Gradient Imbalance  (2610.08315 - Nguyen et al., 6 Oct 2026) in Section 4.4, subsection “Limitations”