Video object detection with a convolutional regression tracker

Video object detection is a fundamental research task for scene understanding. Compared with in images, videos has been less researched due to shortage of labelled video datasets. As frames clip are highly correlated, larger quantity labels needed have good data variation, which not always available as the much more expensive attain. Regarding above-mentioned problem, it easy train an image detector, but possible detector if there insufficient certain classes. In order deal this problem and improve performance classes without labels, we propose augment well-trained efficient effective class-agnostic convolutional regression tracker task. The learns track objects by reusing features from light-weighted increment only slight speed drop our model evaluated on large-scale ImageNet VID dataset. Our strategy improves mean average precision (mAP) score around 5% 3% plus Seq-NMS post-processing.