search

JonnyS1226 / ego4d_asl

code for Ego4D Workshop@CVPR 2023 - 1st in MQ & 2nd in NLQ challenge
9 stars 2 forks source link

Data/Feature download command #1

Closed ardarslan closed 11 months ago

ardarslan commented 11 months ago

Hi,

Thank you for your contribution. Could you please provide the bash command for downloading data and features for the MQ task? I was planning to use the following:

ego4d --output_directory="ego4d_data" --datasets clips annotations omnivore_video_swinl slowfast8x8_r101_k400 --benchmarks mq --version v2

Best regards, Arda

JonnyS1226 commented 11 months ago

Sure. You can download slowfast and omnivore features by the command you mentioned. EgoVLP and InternVideo features can be downloaded from egovlp train&val , egovlp test and internvideo. Combination of features will achieve a better result. We will update instructions in readme later.

ardarslan commented 11 months ago

Thank you for your prompt reply. Is it also true that I should download "clips" and not the "full scale" videos?

JonnyS1226 commented 11 months ago

yes, no need to download full-scale videos if you only want to tackle MQ or NLQ tasks based on this codebase.

ardarslan commented 11 months ago

Thanks a lot once again!

ardarslan commented 11 months ago

This will probably be mentioned in the update in the README but could you please tell how can I construct the features in "/sdb/sjy/projects/ego4d/v1/slowfast_clip" and "/sdb/sjy/projects/ego4d/v1/omnivore_clip"? As far as I understood, the features I downloaded using the bash command above are extracted using untrimmed 30 FPS canonical videos, and for each clip, I should simply crop these features and save them as .pt files in these two directories.

Edit: https://github.com/happyharrycn/actionformer_release/blob/main/tools/convert_ego4d_trainval.py I think this script does what I'm looking for. The problem is solved for now, thank you anyway!