Filtering based on meta data for each youtube URL

[daniel-z-kaplan]

Inappropriate content: NSFW, hate speech, offensive words, sentiments
Title, captions, tags, comments
Quality data:
Length, Aspect Ratio, Resolution, likes, views, number of subscribers of the creator, comments filtered with language model, ensure some amount of movement in the video
Language: English for now

(From Michael) the final version of the dataset from the collection side of things will look like the following:

• channels.tsv with columns ['link', 'name', 'description', 'subscribers', 'isFamilySafe', 'tags']
• videos.tsv with columns ['channel_link', 'id', 'title', 'date', 'length', 'views']
• a sample with a 2 videos from a random set of channels listed in channels.tsv can be found in the DuckAI google drive (note there may be some duplicates!!)

Loose tasks:

• Data analysis on metadata in the TSVs
• Build pipeline to collect metadata using video2dataset (may need to alter video2dataset)
• Analysis on NSFW (using some kind of NSFW filter)
• English filter
• Pixel based filters using thumbnails (see comments below on youtube's auto generated thumbnails)
• Brainstorm more ideas on filtering to have high quality video!

[mtanghu]

We could use VidGear/CamGear to do frame based filtering/streaming (it allow for reading in frames of a video and streaming without needing to download)

[mtanghu]

There are also 4 autogenerated thumbnails from Youtube for each video (from different parts of the video)

https://stackoverflow.com/questions/2068344/how-do-i-get-a-youtube-video-thumbnail-from-the-youtube-api