轨迹质量问题 #24
Description
Hello author, this paper performs quite excellently in video generation. However, in terms of trajectory generation, when reproducing the paper, we found that the metrics for 3 seconds are as follows. They are slightly different from the metrics presented in the paper. Are there any points we need to pay attention to? Additionally, the trajectory output of the NuScenes model is 1.5 seconds, not 3 seconds, right?Freeze all modules, unfreeze the trajectory DIT. The 3-second L2 is approximately around 2.45. The more training epochs, the more severe the trajectory overfitting becomes.
