NTT_CQUPT@TRECVID2019 ActEV: Activities in Extended Video

In this notebook paper, we present our activity detection system, which aims to temporally localize activities in surveillance videos. Our pipeline composed of five modules, object detection, activity proposal generation, feature extraction, classification and post-processing. We input RGB and optical flow into this pipeline separately and obtain frame level predictions by late fusion. The final detections are generated by greedily merging these predictions and filtering invalid results. 1. System Description Activity detection in surveillance videos is a challenging task due to the low resolution, occlusion of objects and similarity between activities. In order to get reliable results, most previous participants used the method of decomposing the task into multiple subtask [1, 2, 3]. Our system for activity detection in extended videos (ActEV) in TRECVID2019[4] is composed of five modules: object detection, activity proposal generation, feature extraction, classification and postprocessing. The diagram of the five modules in our system is shown in Figure 1. We evaluate and analysis each module separately in the following. Object detection: It locates and classifies objects and activities. Activity proposal generation: It generates candidate tubes by temporally tracking bounding boxes for activities. These tubes are called activity proposals. Feature extraction: It finetunes the backbone network and extracts features for activity proposals. Classification: It trains classifier to classify activity proposals. Post-processing: It merges the activity proposals for activity localization. Object Detection Proposal Generation I3D Feature Extraction Classification Postprocessing Input Activity category and location Figure 1: System Overview.