A Multi-Scale Learnable Feature Alignment Network for Video Object Detection

Object detection is an important task of computer vision used to detect instances of visual objects of a certain class in digital images. Video object detection aims to locate single or multiple objects in sequential images and assign category labels for them. There are similarities between video object detection and image object detection, so some image object detection methods are usually used for video object detection. However, due to motion blur, occlusion, morphological diversity, and illumination changes in video, video object detection algorithms have higher requirements. In the framework of video object detection based on feature reuse and recursive fusion, we propose a multi-scale learnable sampling alignment (MLFA) network for video object detection. MLFA divides the video frame into the key frame and non-key frame and propagates a memory feature containing historical key frame information in the time dimension to compensate for the current frame feature through feature fusion. In the process of alignment, the feature pyramid is first established, and then the alignment features of different levels are learned in a learnable way. After that, features from different levels are fused to leverage multi-scale information. MLFA maintains the efficiency and further improves the detection accuracy.