FHO LTA slowfast features missing for some videos

Hello,
I downloaded version 2 for the fho lta (long term anticipation) challenge: features (slowfast) and annotations but in the features some videos are missing.
e.g.

video uid = 44a173ba-e836-4efc-b485-e94d6bbafa8e
video uid = bedaa131-deae-4279-9fd4-5b6ab552644c

don’t have pre-extracted features.

Maybe others are missing but I did not check…
I tried also to download again the features but it says that I up to date with the server.
Can you check if it is only a problem on my side or not?

(PS metadata file does not download automatically if v1 metadata is already present in folder, as stated in doc)

Hi Fede,

Apologies for the missing features. I honestly forgot to extract them out for the release. I can get them for SlowFast within a couple of weeks (currently busy with a lot of items).

Kind regards,
Miguel.

Correction here.

These videos are V1 videos, and therefore do have features extracted for them. For the new videos in V2, I have not exacted features for them, but the videos in FHO/LTA are not new V2 videos.

Unfortunately the metadata for FHO / LTA does not seem to cover all the video uids in the benchmark, hence why when downloading, you have some missing features.

I instead recommend downloading all the features. In other words, remove your --benchmarks filter from the command line, i.e. run:

ego4d  --output_directory=<out-directory> --datasets slowfast8x8_r101_k400

A couple of comments:

  • I do recommend using omnivore video features. If you’re after performance, these are much better features for downstream tasks. Of course if you’re wanting to do an ablation study, ignore this comment.
  • If you’re trying to save space, there are the FP16 omnivore video features available. I believe the total storage size of the FP16 features is around 75GB (don’t quote me on the exact number)

I hope that’s helpful and unblocks you.

2 Likes