CVPR 2022 | querydet: use cascaded sparse query to accelerate small target detection under high resolution

Although in the past few years , Universal target detection based on deep learning has achieved great success , However, the performance and efficiency of detecting small targets are far from satisfactory . The most common and effective method to promote small target detection is to use high-resolution images or feature maps . However , Both methods can lead to expensive calculations , Because the computational cost will increase with the increase of image and feature size .

We proposed QueryDet, it A novel query mechanism is used to speed up the inference speed of target detector based on feature pyramid . The pipeline It consists of two steps : Firstly, the coarse location of small targets is predicted on the low resolution features , Then the accurate detection results are calculated by using the high-resolution features of these coarse position sparse guidance . In this way, high resolution can be obtained feature map Of benefit, It can also avoid using less computation for the background area .

stay popular COCO On dataset , This method will mAP Improved 1.0,mAP-small Improved 2.0, The reasoning speed of high resolution is increased to 3.0×. In a that contains more small objects VisDrone On dataset , We have acquired new SOTA, At the same time, the average 2.3× High resolution acceleration .

Paper title ：QueryDet: Cascaded Sparse Query for Accelerating High-Resolution for Small Object Detection

Thesis link ：https://arxiv.org/abs/2103.09136

Code link ：https://github.com/ChenhongyiYang/QueryDet-PyTorch