Video¶

Continuous stream of images

Before deep learning¶

they used 'tracklets'

Generalizing CNNs¶

Intuitively a recurrent CNNs to share weights across time

Extend convolutional filters in time, performing spatio-temporal convolutions,

so filters will be \(w \times h \times T; T \in [2, 15]\), where \(T =\) temporal width

However, single frame is a very strong baseline and motion detection does not add much value

Comments