![\"display_aspect_ratio\"](\"https://i.imgur.com/8QLANt4.png\")
While looking into sample train videos metadata I found that if display_aspect_ratio field is missing it is a real video, and if it equals 16:9 it is a fake. See the notebook.
Submitting ones and zeros accordingly gave not so good score, which means it is not strictly true in the public test. But it is mainly true. By submitting 0.1 and 0.9 instead I got a gold medal with score 0.54962.
A data leak that needs to be fixed, but not too scary. I think Kaggle needs to\n1. clean metadata from at least public and private test datasets\n2. recalculate the leaderboard
", "messages": [ { "id": "701581", "postDate": "12/23/2019 16:23:58", "content": "While looking into sample train videos metadata I found that if display_aspect_ratio field is missing it is a real video, and if it equals 16:9 it is a fake. See the notebook.
Submitting ones and zeros accordingly gave not so good score, which means it is not strictly true in the public test. But it is mainly true. By submitting 0.1 and 0.9 instead I got a gold medal with score 0.54962.
A data leak that needs to be fixed, but not too scary. I think Kaggle needs to\n1. clean metadata from at least public and private test datasets\n2. recalculate the leaderboard
", "rawMarkdown": "While looking into sample train videos metadata I found that if `display_aspect_ratio` field is missing it is a real video, and if it equals `16:9` it is a fake. See [the notebook](https://www.kaggle.com/zaharch/data-leak-in-metadata).\n\n\n\nSubmitting ones and zeros accordingly gave not so good score, which means it is not strictly true in the public test. But it is mainly true. By submitting `0.1` and `0.9` instead I got a gold medal with score `0.54962`.\n\nA data leak that needs to be fixed, but not too scary. I think Kaggle needs to\n1. clean metadata from at least public and private test datasets\n2. recalculate the leaderboard", "votes": null }, { "id": "701604", "postDate": "12/23/2019 16:46:30", "content": "Hi @zaharch - thanks! This was pointed out to us over the weekend by @jarfo1, and we're in the process of correcting it, using exactly that approach. I'll post a thread when we're going to start updating things.
", "rawMarkdown": "Hi @zaharch - thanks! This was pointed out to us over the weekend by @jarfo1, and we're in the process of correcting it, using exactly that approach. I'll post a thread when we're going to start updating things.", "votes": null }, { "id": "701671", "postDate": "12/23/2019 18:50:39", "content": "same is true for sample_aspect_ratio. 0:1 correlates to real, 1:1 correlates to fakes. they map with display_aspect_ratio
Yep! This was also noted by @jarfo1 and we're working on it as well. Thank you!
", "rawMarkdown": "Yep! This was also noted by @jarfo1 and we're working on it as well. Thank you!", "votes": null }, { "id": "701717", "postDate": "12/23/2019 20:40:38", "content": "@philculliton , hope the changes are going to be only in the validation & test sets and not in the training set. Else, we will be forced to re-download the training set. Pls confirm.
", "rawMarkdown": "philculliton , hope the changes are going to be only in the validation & test sets and not in the training set. Else, we will be forced to re-download the training set. Pls confirm.", "votes": null }, { "id": "701719", "postDate": "12/23/2019 20:47:43", "content": "@ravivadapalli - That's correct! The plan is only to change the validation and test sets as necessary. Should be no need to re-download the training set.
", "rawMarkdown": "ravivadapalli - That's correct! The plan is only to change the validation and test sets as necessary. Should be no need to re-download the training set.", "votes": null }, { "id": "703741", "postDate": "12/26/2019 15:35:21", "content": "Good catch!
", "rawMarkdown": "Good catch!", "votes": null }, { "id": "747763", "postDate": "02/16/2020 20:15:31", "content": "Good observation! Any pointer if this has been fixed?
", "rawMarkdown": "Good observation! Any pointer if this has been fixed?", "votes": null }, { "id": "747781", "postDate": "02/16/2020 20:56:37", "content": "Yes, it's been fixed. Check this out.
", "rawMarkdown": "Yes, it's been fixed. Check [this](https://www.kaggle.com/c/deepfake-detection-challenge/discussion/124256) out.", "votes": null }, { "id": "747792", "postDate": "02/16/2020 21:27:58", "content": "Thanks for the link.
", "rawMarkdown": "Thanks for the link.", "votes": null } ], "comments": [ { "id": 701604, "author_name": "philculliton", "author_url": "", "post_date": "12/23/2019 16:46:30", "content": "Hi @zaharch - thanks! This was pointed out to us over the weekend by @jarfo1, and we're in the process of correcting it, using exactly that approach. I'll post a thread when we're going to start updating things.
", "votes": null, "replies": [ { "id": 701717, "author_name": "ravivadapalli", "author_url": "", "post_date": "12/23/2019 20:40:38", "content": "@philculliton , hope the changes are going to be only in the validation & test sets and not in the training set. Else, we will be forced to re-download the training set. Pls confirm.
", "votes": null, "replies": [] }, { "id": 701719, "author_name": "philculliton", "author_url": "", "post_date": "12/23/2019 20:47:43", "content": "@ravivadapalli - That's correct! The plan is only to change the validation and test sets as necessary. Should be no need to re-download the training set.
", "votes": null, "replies": [] } ] }, { "id": 701671, "author_name": "tivfrvqhs5", "author_url": "", "post_date": "12/23/2019 18:50:39", "content": "same is true for sample_aspect_ratio. 0:1 correlates to real, 1:1 correlates to fakes. they map with display_aspect_ratio
Yep! This was also noted by @jarfo1 and we're working on it as well. Thank you!
", "votes": null, "replies": [] } ] }, { "id": 703741, "author_name": "liuhh02", "author_url": "", "post_date": "12/26/2019 15:35:21", "content": "Good catch!
", "votes": null, "replies": [] }, { "id": 747763, "author_name": "yassinealouini", "author_url": "", "post_date": "02/16/2020 20:15:31", "content": "Good observation! Any pointer if this has been fixed?
", "votes": null, "replies": [ { "id": 747781, "author_name": "msheriey", "author_url": "", "post_date": "02/16/2020 20:56:37", "content": "Yes, it's been fixed. Check this out.
", "votes": null, "replies": [] }, { "id": 747792, "author_name": "yassinealouini", "author_url": "", "post_date": "02/16/2020 21:27:58", "content": "Thanks for the link.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "701581": "While looking into sample train videos metadata I found that if `display_aspect_ratio` field is missing it is a real video, and if it equals `16:9` it is a fake. See [the notebook](https://www.kaggle.com/zaharch/data-leak-in-metadata).\n\n\n\nSubmitting ones and zeros accordingly gave not so good score, which means it is not strictly true in the public test. But it is mainly true. By submitting `0.1` and `0.9` instead I got a gold medal with score `0.54962`.\n\nA data leak that needs to be fixed, but not too scary. I think Kaggle needs to\n1. clean metadata from at least public and private test datasets\n2. recalculate the leaderboard", "701604": "Hi @zaharch - thanks! This was pointed out to us over the weekend by @jarfo1, and we're in the process of correcting it, using exactly that approach. I'll post a thread when we're going to start updating things.", "701671": "same is true for `sample_aspect_ratio`. `0:1` correlates to real, `1:1` correlates to fakes. they map with `display_aspect_ratio`", "701681": "Yep! This was also noted by @jarfo1 and we're working on it as well. Thank you!", "701717": "philculliton , hope the changes are going to be only in the validation & test sets and not in the training set. Else, we will be forced to re-download the training set. Pls confirm.", "701719": "ravivadapalli - That's correct! The plan is only to change the validation and test sets as necessary. Should be no need to re-download the training set.", "703741": "Good catch!", "747763": "Good observation! Any pointer if this has been fixed?", "747781": "Yes, it's been fixed. Check [this](https://www.kaggle.com/c/deepfake-detection-challenge/discussion/124256) out.", "747792": "Thanks for the link." }, "source": "meta" }