[iccv workshop 2021] small target detection based on density map: coarse-grained density map guided object detection in aerial images

No open source code found

brief introduction ：

Object detection in aerial images is challenging , There are at least two reasons ：（1） Relative to high-resolution aerial images , Most targets are small-scale ;（2） The target position is unevenly distributed , The detection efficiency is low .

In view of the above two problems , This paper presents a new network , That is, coarse-grained density graph network （CDMNet） To solve these problems . say concretely , The author formats the density mapping into a coarse-grained form , A lightweight dual task density estimation network is designed .

The coarse-grained density graph can not only describe the distribution of objects , You can also cluster objects , In order to quantify the scale and reduce the calculation .

Besides , The author also proposes a clustering region generation algorithm guided by density mapping , Crop the input image into multiple sub regions , Represents clustering , The object is adjusted on a reasonable scale .

The author also improved the mosaic data enhancement , In order to alleviate the imbalance of foreground background and category in the training process .