Improving real-time small objects detection by fusion features of spatial coordinates

The small object detection has long limited the advanced development of deep learning-based detection models. The down-sampling of vision-models leads to a severe loss of spatial information of small objects causing a great increase in the difficulty of models to capture the location of small ones. Inspired by the logic of human vision perception behavior, we found that the main limitation of small object detection is the location regression rather than classes differentiation. Therefore, we first introduce coordinate features to perform multi-scale spatial information perception and element-level width-height independent coordinate encoding of image features in anticipation of easing the difficulty of small-object detection. Secondly, we design a lightweight architecture, Coord-Yolo-E, for the real-time small-object detection based on the coordinate feature scheme. Finally, we accomplish 23.1 and 40.6% AP on VisDrone and MS COCO benchmark by the Coord-Yolo-E within 7.9 million parameters, which provides a superior accuracy-speed trade-off compared to state-of-the-art one-stage real-time detectors.