Some fho oscc videos not under fho benchmark

Q&A
1

Hello,

I am trying to download the full_scale videos of all the video_uids with fho oscc annotations. I gathered the video uids in a list by reading the fho_oscc-pnr_{stage}.json files and passed them to the Ego4D CLI in a command like so:

ego4d -o /data/Ego4D/ --datasets full_scale --benchmarks FHO --version v2 --video_uid_file /data/fho_oscc_v2_video_uids.txt

This quickly returns the error:

RuntimeError: The following requested video UIDs could not be found in the manifest for version: ‘v2’ and dataset(s): ‘{‘full_scale’}’:
{‘video_uid1’, ‘video_uid2’, … , ‘video_uid38’}

And then lists just 38 video uids.

Interestingly, if I DON’T include the --benchmarks FHO flag, it proceeds just fine and prompts me to continue downloading. Why might this be the case? The specified 38 video uids seem to be listed as FHO in the Ego4D visualizer.

Some example uids from the 38 listed in the error message: e8c52c5d-4d8f-4d9f-befd-29f661f80f27, 98d8c17a-ce41-4b9d-ae28-ca785d60be01, 0a90656a-eac8-49fa-a253-78f5626c8d5c

2

Hi,

Thank you for bringing this to our attention. The metadata for the --benchmark FHO filter seems to not be complete (either due to a bug or otherwise). I am currently investigating this and will post an update here when this is fixed.

The reason you are seeing the CLI downloading the videos without the --benchmarks FHO flags is because of the way filtering works. Essentially every filter flag (--benchmarks and --video_uid_file) defines a set of video uids to consider, and the resulting video uid set is defined by the AND of these sets. Hence, because not all videos are flagged as being in FHO, this is why you notice that excluding --benchmarks FHO works properly.

Thanks for your understanding,
Miguel.

1 Like