It is not clear to me if the method developed for this challenge is required to \nto detect whether a suspect video is a deepfake by processing one video at a time or not.\nRight now the public test allows seeing all the videos at the same time, \nhence it is possible to adopt a solution based on pseudo-labeling.\nIs this allowed?
\nMore in general, is it allowed to exploit all videos of the test set (private leaderboard) \nto improve performance on each single one of them?
It is not clear to me if the method developed for this challenge is required to \nto detect whether a suspect video is a deepfake by processing one video at a time or not.\nRight now the public test allows seeing all the videos at the same time, \nhence it is possible to adopt a solution based on pseudo-labeling.\nIs this allowed?
\nMore in general, is it allowed to exploit all videos of the test set (private leaderboard) \nto improve performance on each single one of them?
Yes, it is allowed, you are given test_videos folder and you can process it in whatever order and way you want, as long as you produce submission.csv at the end.
May not be a good idea because pseudo labeling increases risk of overfitting.
", "rawMarkdown": "May not be a good idea because pseudo labeling increases risk of overfitting.", "votes": null }, { "id": "748032", "postDate": "02/17/2020 06:05:15", "content": "Ideally, you should process one video at a time and assume nothing about the other videos in the test set; test yourself in the extreme case where every video comes from a different actor and every video is unique. If anyone wants to explore other solutions based on pseudo-labelling, keep in mind that overfitting may occur (as someone already mentioned).
", "rawMarkdown": "Ideally, you should process one video at a time and assume nothing about the other videos in the test set; test yourself in the extreme case where every video comes from a different actor and every video is unique. If anyone wants to explore other solutions based on pseudo-labelling, keep in mind that overfitting may occur (as someone already mentioned).", "votes": null } ], "comments": [ { "id": 735723, "author_name": "zaharch", "author_url": "", "post_date": "02/03/2020 11:39:14", "content": "Yes, it is allowed, you are given test_videos folder and you can process it in whatever order and way you want, as long as you produce submission.csv at the end.
May not be a good idea because pseudo labeling increases risk of overfitting.
", "votes": null, "replies": [] }, { "id": 748032, "author_name": "cristiancanton", "author_url": "", "post_date": "02/17/2020 06:05:15", "content": "Ideally, you should process one video at a time and assume nothing about the other videos in the test set; test yourself in the extreme case where every video comes from a different actor and every video is unique. If anyone wants to explore other solutions based on pseudo-labelling, keep in mind that overfitting may occur (as someone already mentioned).
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "735717": "It is not clear to me if the method developed for this challenge is required to \nto detect whether a suspect video is a deepfake by processing one video at a time or not.\nRight now the public test allows seeing all the videos at the same time, \nhence it is possible to adopt a solution based on pseudo-labeling.\nIs this allowed? \nMore in general, is it allowed to exploit all videos of the test set (private leaderboard) \nto improve performance on each single one of them?", "735723": "Yes, it is allowed, you are given `test_videos` folder and you can process it in whatever order and way you want, as long as you produce `submission.csv` at the end.", "735750": "May not be a good idea because pseudo labeling increases risk of overfitting.", "748032": "Ideally, you should process one video at a time and assume nothing about the other videos in the test set; test yourself in the extreme case where every video comes from a different actor and every video is unique. If anyone wants to explore other solutions based on pseudo-labelling, keep in mind that overfitting may occur (as someone already mentioned)." }, "source": "meta" }