Answers are grounded in this paper and cite their sources.
Vehicle detection in aerial imagery suers from small scales, arbitrary orientations, and limited single-modality information. Multi-modal RGBIR fusion is widely adopted, yet the eld rests on two tacit assumptions rarely veried: (1) paired RGB and IR images are spatially misaligned, making alignment the main obstacle to fusion; and (2) IR assists ne-grained class discrimination. Using a YOLOv8s-based rotated-box multi-modal detector , we reexamine both assumptions on VEDAI. Phase-correlation measurements show a median misalignment of only 0.078 pixels, indicating the dataset is well registered; permutation tests show IR adds no signicant discriminative power for the confusable carpickup pair (p = 0.81). Data proling instead reveals the true challenge: class confusion among extremely small objects (95.3% smaller than 32×32 pixels), primarily geometric. The framework nonetheless yields consistent gains: the rotated-box head improves mAP50-95 by 12.86 points and IR adds 5.79 points (paired t-test p < 0.0001 and p = 0.0026). An explicit alignment module (BMA) brings no signicant gain on these well-registered datasets (eight seeds; equivalence test and power analysis; mean +0.44 mAP50, p = 0.563). Reproducing this on DroneVehicle, alignment again yields no reliable gain (+0.21 mAP50-95), indicating the null holds across two independently registered datasets; the misaligned regime is tested only synthetically.
No takes yet. Share an insight, caveat, or question.
Yi Wu (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: