Timed-image versus dynamic-image representations of a video sequence

From Timed-Image Based Deep Learning for Action Recognition in Video Sequences, by A.M. Atto, A. Benoit and P. Lambert.

For the timed-image representation, the N-th column of the image given at the right-hand-side of the table corresponds to the N-th frame of the video sequence displayed at left-hand side (any frame is reshaped with respect to a Hilbert space filling curve for locality preservation). This implies no information loss, in contrast with an optical flow-based image integration (dynamic-image).

Example of video clips (top-left), their corresponding Timed-images (right) and still-dynamic images (down-left)