Reproduce K400 with Video-FocalNet-T about video-focalnets HOT 2 CLOSED

Hi,

Thank you for reaching out.

Please note that we report the results on the Kinetics-400 Validation set, following the common practice of other popular works in literature, where they train on the training set and report on the validation set.

We also provide the log file below to show the results. Additionally, if required, we can make available our resized version of the Kinetics-400 train and validation sets. Please let us know if you would like us to do that.

log_rank0.txt.

Kind regards,

from video-focalnets.

@TalalWasim
Thanks for your detail response.

I used K400 test set, link. It's collected from open-datalab. Hopefully it's much as same as yours. The version that you mentioned, if possible please share.

Thanks for sharing the log file, it's helpful. The mian pionts probably are as follows for inference:

INPUT_SIZE: 224
NUM_FRAMES: 8
TEST:
  CROP: true
  NUM_CLIP: 4
  NUM_CROP: 3

For now, I didn't use multiple crop, and use only single spatial view. I think not using 3 crop will not cause major score gap.

In the paper, page 7, it is mentioned as below:

During inference, we report results as an average across Nclip × Ncrops
where a total of Nclip clips are uniformly sampled from the video, and for each video, Ncrops spatial crops are taken during inference.

To follow up, I also used n-clip=3 and iterate inference on multiple clip and average next for a single instance. By doing so, scores improves 65% to 71%.

update

reproduced. great. For me, It's was about sampling.

from video-focalnets.