Temporal ROI Align for Video Object Recognition

Video object detection is challenging in the presence of appearance deterioration certain video frames. Therefore, it a natural choice to aggregate temporal information from other frames same into current frame. However, ROI Align, as one most core procedures detectors, still remains extracting features single-frame feature map for proposals, making extracted lack videos. In this work, considering instance are highly similar among video, novel Temporal Align operator proposed extract maps frame proposals by utilizing similarity. The can entire proposals. We integrate detectors and state-of-the-art conduct quantitative experiments demonstrate that consistently significantly boost performance. Besides, also be applied segmentation.