CARLA Town01

Description

Example completions, starting from the first 36 frames of a test video. Within each block each column shows completions for a different test video. The top row is the ground-truth and the bottom row shows sampled completions. Observed frames are marked with a red border, and the end of the video is marked with a checkerboard pattern.

Discussion

Samples from FDM with Autoreg show no movement a large proportion of the time. This is a similar failure mode to that shown for the Long-range sampling scheme in Figure 6, and explains the bad WD score in Table 1. Samples from FDM with Hierarchy-2 exhibit occasional "flickers", where frames generated in the first stage of the sampling scheme are not aligned with neighbouring frames generated in later stages (explaining the relatively bad OP score). They are otherwise fairly realistic (explaining their relatively good WD and FVD scores). As seen in the other datasets, samples from CWVAE are much blurrier than those from the other methods. Samples from VDM do not show obvious failure modes, but an analysis of the point-speeds of VDM's generated videos (as described in Section 5) shows that the speed is excessively variable. VDM point-speeds have smaller peaks around zero and the 3m/s speed limit than the ground-truth, FDM with Autoreg, or FDM with Hierarchy-2.

FDM with Autoreg

FDM with Hierarchy-2

CWVAE

VDM