Multi-level Proposal Relations Aggregation for Video Object Detection

Video information often deteriorates in certain frames, which is a great challenge for object detection. It is difficult to identify the object in this frame by just utilizing the information of one frame. Recently, plenty of studies have shown that context aggregating information through the self-attention mechanism can enhance the features in key frames. However, these methods only exploit some of inter-video and intra-video global-local information, not all of it. Global semantic and local localization information in the same video can assist object classification and regression. The intra-proposal relation among different videos can provide important cues to distinguish confusing objects. All of this information is able to enhance the performance of video object detection. In this paper, we design a Multi-Level Proposal Relations Aggregation network to mine inter-video and intra-video global-local pro-posal relations. For intra-video, we effectively aggregate global and local information to augments the proposal features of key frames. For inter-video, we aggregate the inter-video key frame features to the target video under the constraint of relation regularization. We flexibly utilize the relation module to aggregate the proposals from different frames. Experiments on ImageNet VID dataset demonstrate the effectiveness of our method.