S-Feature Pyramid Network and Attention Module For Small Object Detection

Because of the low resolution and limited information of small objects, and the computing resources are limited in practical applications, small object detection is still challenging. In order to improve the accuracy of small object detection, we propose a new method. It’s included a shallow feature pyramid network with an information extraction block at the shallow features and fused multi-scale semantic information. Further, context information with attention mechanism is adopted to make object detection focus on the significant area. We are one of the top five teams in the Drone-vs-Bird Detection Grand Challenge. The detection ability of our method for small objects is much higher than classical one-stage and two-stage detectors. For limited computer resources, 300×300 inputs are used and the detection speed of 45 fps is reached by the proposed method, which can realize real-time object detection.

1. Introduction

With the development of deep convolution networks, the performance of object detection has been significantly improved. However, small object detection is still a challenging research, as there is less information in the area of the small object in the image and low resolution. According to MS COCO [1] definition, define object with less than 32× 32 pixels as small objects. Many advanced object detectors can perform well in their training dataset, but it lacks of robustness to small objects in actual scenes. And its advanced performance is due to a large number of parameters and larger pixel input images. Therefore, the existing small object detection methods can’t be well deployed in the application scenarios with limited computing resources, and it is difficult to meet the requirements of real-time detection. There is still a lot of room to improve in small object detection.