Pertinence of Video for Single Image Deep Network

摘要

Using key frames instead of video to train single image deep neural networks make sense as successive images of one video contain almost the same information. However, we show that using all images can significantly increase performances of deep networks on medium size datasets. Considering, that annotating video can be done much more efficiently than annotating disparate images, we argue that using complete videos should be considered where data are naturally collected this way which is often the case in robotic, autonomous driving, or aerial acquisitions.