search

Araachie / river

Efficient Video Prediction via Sparsely Conditioned Flow Matching. In ICCV, 2023.
https://araachie.github.io/river
GNU General Public License v3.0
8 stars 1 forks source link

Evaluation Metrics #5

Open emmahart2000 opened 1 month ago

emmahart2000 commented 1 month ago

How were the metrics you present in the paper computed?
For example, I am trying to recreate your KTH results. How were the FVD, PSNR, and SSIM computed across the different actions?

Thank you, all the best

Araachie commented 1 month ago

Hi,

Thanks for the interest in our paper and code!

For the evaluation, we first generated videos starting from a number of initial ground truth frames from the test set (for KTH we generated 30 or 40 frames from initial 10; for the details on the train/test split check this repo). Then we used the generated and the ground truth videos to calculate the metrics. For this, we used the evaluation scripts from this repo. Note that our method is unconditional, so no action information was used. We reported the average PSNR and SSIM across generated frames.