Non-dense Feature Aggregation for Video Object Detection

Abstract

Video object detection is a challenging task in computer vision. Dynamic background change, motion blur, motion occlusion, and other problems bring great interference to the detection results. The existing detection methods often sacrifice the detection speed for the sake of accuracy or sacrifice the accuracy for the sake of faster detection speed. Therefore, we propose a non-dense feature aggregation method based on an optical flow network. The feature extraction is only carried out for keyframes, and then the feature aggregation of non-key frames is completed by the non-dense connection of adjacent frames, and the idea of dynamic programming is used to achieve the transmission of historical information, so as to obtain more effective features for detection. Our algorithm achieves a real-time detection effect without significantly increasing the amount of computation. And the algorithm achieves competitive detection results on the ImageNet VID dataset.