Dreamer Performance During Aligning and Co-Training Stages

[suhaisheng]

How about the Top1 accuracy of gen mean and gen tail across 10 scales of the goal image in both stages? And what is the interval of sampled frame?

[aopolin-lv]

Thanks for your interest.

1. Top1 acc
   The top-1 accuracy of gen mean and gen tail is approximately 40-50%.

2. Interval of sampled frame
   During Pretraining Stage I, we uniformly sample up to 80 frames from each video. In Stages II and III, however, the sampling interval varies across subsets. The key principle is to maintain an effective temporal spacing of roughly 0.5 seconds between sampled frames.