Cheaper Clicks from Boxes: Cyclic Querying for Interactive Small Object Detection

Interactive object detection paves a promising way to alleviate the annotation burden, particularly for size-limited instances in remote sensing images. While early studies emphasize fortifying object representations through click-based interactions, the alternative strategy of harnessing image as a prior to derive informative clicks remains largely under-explored - despite its substantial potential in advancing interactive understanding. To address this limitation, we devise a CycLIc querying (CLIQ) paradigm tailored for interactive small object detection in remote sensing imagery. The core lies in an interactive semantic querying framework, where user clicks and visual representations interact in a mutually reinforcing fashion: the click prior first shepherds feature modulation and correlates with a set of classwise queries to elicit rich semantic responses; the resulting heatmap, in turn, dynamically functions as a spatial activator to highlight small-instance regions. Furthermore, we devise a simple yet effective Informative Click Querying strategy to pinpoint potential click candidates, thereby completing the closed image-to-click interaction loop and ameliorating the performance under few-click scenarios. Thanks to its lightweight CLIQ design, the proposed CLIQ framework achieves competitive results on the AITOD-R and Tiny-DOTA benchmarks with negligible computational overhead, significantly outperforming pioneering attempts in small object detection. More importantly, the superior performance of our interactive paradigm under low-data regimes underscores its strong potential as a cost-effective and adaptive alternative for data-efficient detection workflows. The code is available at: https://github.com/shaunyuan22/CLIQ.