Differences between PQ, SQ and RQ obtained during the training process and those obtained during the evaluation process

[Francisco-liu]

Hello, I found that the metrics in the training script train_panopli_tensorf.py and the evaluation script evaluate.py are calculated differently. There is a big difference between the PQ, SQ, and RQ obtained during the training process and the evaluation process. why is that?

[nihalsid]

Hi,

The ones during training are just a rough estimate, since the evaluation is too expensive to do during the training.

Regards, Yawar Nihal

On Sun 19. May 2024 at 11:16, Francisco-liu @.***> wrote:

Hello, I found that the metrics in the training script train_panopli_tensorf.py and the evaluation script evaluate.py are calculated differently. There is a big difference between the PQ, SQ, and RQ obtained during the training process and the evaluation process. why is that?

— Reply to this email directly, view it on GitHub https://github.com/nihalsid/panoptic-lifting/issues/43, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAHDSDQ2D3P6PMUKRA2HLHTZDB3XRAVCNFSM6AAAAABH6HWGNOVHI2DSMVQWIX3LMV43ASLTON2WKOZSGMYDINJTGY3DINI . You are receiving this because you are subscribed to this thread.Message ID: @.***>

[Francisco-liu]

Is it too expensive to evaluate during training because of step_ratio? i.e.renderer.update_step_ratio(renderer.step_ratio * 0.5). However, I found that the PQ, SQ, and RQ during the training process are better than those during the evaluation process.