Existing object detection methods primarily focus on detecting large targets in images, with limited research conducted on small targets. Additionally, there are challenges with low detection accuracy and difficulty in meeting real-time requirements. Based on this, an improved RetinaNet is proposed. First, the use of high-resolution feature maps can better capture the features of small targets. Then a weight factor is added to solve the unbalanced distribution of training samples caused by using high-resolution feature maps. Finally, adding a novel query subnet by two main steps: first estimate the coarse location of the small target in the high-level feature map, then the exact location is calculated in the low-level feature map. Experiments show that this query subnet saves the computational cost of low-level feature maps, and greatly improved detection speed. The performance of the model is evaluated on a new benchmark called TinyPerson, mAP up to 36.18, FPS up to 13.98. Compared to the unimproved RetinaNet, the accuracy and speed are improved by 0.9 and 1.28 respectively.
|