Rethinking Focus-and-Detect for Aerial Small Object Detection Under a Frozen-Detector Protocol: A Translation-Equivariance-Guided Approach

Objects in unmanned aerial vehicle (UAV) imagery are typically small, densely packed, and unevenly distributed. The prevailing focus-and-detect paradigm—first localizing focus regions where objects concentrate, and then performing detection within them—has substantially improved small object detection. Existing focusing modules, however, are almost always coupled with architectural modifications to the detector and with patch-based training, which makes the contribution of the focusing mechanism itself difficult to quantify. More importantly, whether the focusing mechanism alone can yield genuine gains on off-the-shelf detectors without any retraining remains insufficiently examined. To address these issues, we first integrate the focusing modules of several representative open-source methods into frozen detectors, removing the architectural modifications and stripping away the patch-based training on which these modules rely, so as to isolate the genuine gain of the focusing mechanism. We then establish translation equivariance as a design principle for focus-region generation and propose a strictly translation-equivariant heatmap network for this purpose. The resulting focus regions are further processed via super-resolution and image packing to construct focus images that are better suited to frozen detectors. Extensive experiments on VisDrone show that, with a strong baseline, directly plugging existing focusing modules into a frozen detector generally hurts rather than helps; when the baseline is weakened by lowering the input resolution, their gains reappear, but the final accuracy remains substantially below that of the high-resolution baseline. In contrast, our method raises the mean average precision (mAP@0.5:0.95) of multiple strong baseline detectors by 0.44–1.03 points on VisDrone-test-dev, without any retraining. When further combined with patch-based training, the overall framework achieves a total gain of 1.24–2.22 points over the strong baseline.