A Video Object Detector Based on Deep Reinforcement Learning and Correction Module

As a basic research topic of computer vision, object detection is facing the challenges of being extended for video object detection. For deformation anomaly, low-quality candidate regions feature can be caused by motion blur, rare poses and video defocus, resulting to poor detection performance in videos. In this paper, we propose a video object detector, which consists of an image-based detector using deep reinforcement learning and a correction module using objects tracking algorithms. We treat videos as continuous images. The agent will search on the feature map of each image by a policy network to generate high-quality candidate regions. These candidate regions will be corrected by a correction module. The bounding box at the current image from the tracklet proposed by the object tracking algorithms will correct the candidate regions. Experiments are conducted on the MOT 15 dataset. Our detector achieves better detection performance, exactly 68.28% with 5.01 points improvement compared to the Faster-R CNN as the single-frame baseline.