I was wondering, with the advent of deep learning, many tasks related to images have been solved to near human accuracy such as classification, object detection etc., however in videos, traditional methods still compare to deep learning methods and both are not as good. I wanted to understand what and how is it that with videos that understanding them makes it much harder.
I've used py-faster-rcnn on video feeds and found it to be pretty usable in a sunny environment. Although on a GTX660, I could only process one image every 600ms, which may highlight a limitation. It still takes a lot of time to compute an inference on a deep net, so for a minimally fast framerate, of 10fps, you're still limited to a speed of 100ms.
There might also be other artifacts due to camera dynamics such as image blurring, sensor saturation, or low light levels that might be more present in video feeds of natural scenes as opposed to the fairly well exposed images in static image datasets.