[iccv 2019] characteristics precise supervision of feature super resolution for small object detection

Abstract ：

Despite recent proposals based on CNN The model has achieved success in target detection , But because of the small ROI The information contained in the region is limited and distorted , Detecting small targets is still difficult .

One way to alleviate this problem is to use super-resolution technology to enhance small ROI Characteristics of , Therefore, the author studies how to improve the feature level super-resolution, especially for small object detection , And found that its performance can be significantly improved .

To enrich the information in the tip , Some previous studies have used image super-resolution . However , The existing feature level super-resolution models for small target detection have a significant limitation ： Lack of direct supervision . in other words , Their super-resolution model does not train clear target characteristics , This leads to the instability of training and limits the quality of super-resolution features .

Therefore, a new feature level super-resolution method is proposed , Using appropriate high-resolution target features as the training method of surveillance signal SR The model and the matched relative receptive field are trained to input low-resolution features and target high-resolution features , And it can be integrated with any detector based on feature pooling .

The contributions of this paper are as follows ：

• The existing feature level super-resolution methods for small target detection are thoroughly examined , Discover and utilize high-resolution target features as monitoring