Cross-family validation of FD-CanKD

Validate the FD-CanKD alignment principles on detector families beyond the YOLOv12 teacher–student family, including DETR-style detectors and other YOLO variants, to determine whether the observed relation-aware and frequency-aware distillation benefits generalize across architectures.

Background

The study deliberately restricts its experiments to a controlled YOLOv12-X teacher and YOLOv12-M student setting so that the effects of cross-attention-based relation transfer and frequency-decoupled alignment can be isolated from differences in detector architecture, codebase, training schedule, and initialization. Consequently, the reported results do not establish whether FD-CanKD generalizes to other detector families.

The paper explicitly identifies validation on DETR-style detectors and other YOLO variants as an unresolved extension. Establishing such generalization would test whether the proposed framework is genuinely detector-family independent or whether its effectiveness depends on the architectural and feature-level correspondence available within YOLOv12.

References

We note that all reported results are obtained within the YOLOv12 teacher--student family. This single-family control was a deliberate design choice for isolating component effects, and validating the proposed alignment principles on other detector families (e.g., DETR-style or other YOLO variants) is left for future work.

— FD-CanKD: Frequency-Decoupled Cross-Attention Distillation as a Refinement Prior for Compact Object Detectors  (2608.18590 - Cheong et al., 19 Aug 2026) in Discussion section