summarized by : Anonymous
QueryDet: Cascaded Sparse Query for Accelerating High-Resolution Small Object Detection


effectively detecting small objects using high-resolution feature maps (e.g., P2 in FPN) for feature pyramid-based object detectors by Cascade Sparse Query (CSQ); CSQ reduces computation cost


making predictions only on locations potentially containing small objects (Sparse Query, identified by small object existence prediction on coarse feature maps) on high-resolution feature maps


better performance obtained compared with RetinaNet and FOCS on COCO, VisDrone


training of classification and regression branches is kept the same as in the original RetinaNet; a query head for small object prediction (classification task, not regression) is added