Short-term anchor linking and long-term self-guided attention for video object detection

We present a new network architecture able to take advantage of spatio-temporal information available in videos boost object detection precision. First, box features are associated and aggregated by linking proposals that come from the same anchor nearby frames. Then, we design attention module aggregates short-term enhanced exploit long-term information. This takes geometrical for first time video domain. Finally, double head is fed with both spatial reference frame into account short- temporal context. have tested our proposal five datasets very different characteristics, order prove its robustness wide number scenarios. Non-parametric statistical tests show approach outperforms state-of-the-art. Our code at https://github.com/daniel-cores/SLTnet.