We include some videos showcasing our results on VidOR in the supplementary material. For clarity, each video file only shows the predicted relation instances with high scores. The subject is with red boxes, while the object is with blue boxes. In many cases, our approach detects the relation successfully, e.g. 3383804222.mp4 and 3830567237.mp4. For long-range relations such as chase and clean, our method can detect the correct interactions. Some failure cases are due to the misrecognized object (e.g. 8396167542.mp4). Some failure cases are due to the inaccurate and redundant object bounding boxes (e.g. 8826744002.mp4).