Planning KL Calculation


I am trying to compute the planning KL divergence value for some models. I can see this in the output for the CenterPoint model:
{‘min’: 0.017909612506628036, ‘max’: 245.0158233642578, ‘mean’: 13.34814739227295, ‘median’: 4.651473045349121, ‘std’: 25.597375869750977}
On the leaderboard, I see the values are in the range of 0.5-1.0 for all configurations of CenterPoint. Could you provide details on what I am missing here.

Hi @Shubham_Gupta, we provide median PKL scores in the leaderboard. During the PKL score computation, we are only using the predicted boxes with confidence scores higher than optimal scores at max F1. So there are much fewer low-confidence boxes, which leads to better PKL scores.

1 Like

Hi @Yiluan,
Thanks a lot for your response. I would like to know if this max F1 score is computed on the training set or validation set. Thanks

Hi @Shubham_Gupta, it is the validation set.

1 Like

Hi @Yiluan, thanks a lot for providing clarification.

1 Like