Adaptive multi-scale feature aggregation for video object detection

Abstract

We address an efficient object detection framework for videos. Despite high performance in many object detection methods using deep learning methods, there exist some cases that lower the detection performance in videos, such as blur due to the fast movement of objects or camera, occlusion, rare pose, etc. In this paper, to solve the above challenging problems, we propose an adaptive multi-scale feature aggregation method and design a new network for the proposal. Based on the one-stage object detection framework, we aggregate several adjacent frames' features in multi-scale to make it more robust on object size and learn the adaptive weights for the aggregation depend on the quality of features. We show that our proposed method can learn the adaptive weights throughout the network and can improve the performance of video object detection in the feature aggregation stage.