Long time ago I realized that by a simple frame difference rmse clustering, one can detect pairs of (true, fake) videos that are from the same original video. By that way, we can simply correct predictions of each pair. I guess this is not what the organizers want, because it's tricky.
\n\nAs a result, does the private test set has that property? If yes, then the competition solution will be tricky, and I hope the organizers will remove it. Personally, I hope it does not have that property.
\n\nPlease clarify @cristiancanton
", "messages": [ { "id": 791815, "postDate": "2020-03-30T17:27:46.220Z", "content": "Not a 100% proof, but I checked by LB probing whether there are identical audio hashes in the public test, and there are none. All audio hashes are different. It (almost) proofs that the public test dataset does not contain such pairs.
", "rawMarkdown": "Not a 100% proof, but I checked by LB probing whether there are identical audio hashes in the public test, and there are none. All audio hashes are different. It (almost) proofs that the public test dataset does not contain such pairs.", "votes": 3 }, { "id": 791776, "postDate": "2020-03-30T16:43:51.277Z", "content": "nice idea! But obviously it's cheat and does not worth money because this trick does not decide problem of detecting fake faces in real life.
", "rawMarkdown": "nice idea! But obviously it's cheat and does not worth money because this trick does not decide problem of detecting fake faces in real life. ", "votes": 1 }, { "id": 791756, "postDate": "2020-03-30T16:28:32.680Z", "content": "Long time ago I realized that by a simple frame difference rmse clustering, one can detect pairs of (true, fake) videos that are from the same original video. By that way, we can simply correct predictions of each pair. I guess this is not what the organizers want, because it's tricky.
\n\nAs a result, does the private test set has that property? If yes, then the competition solution will be tricky, and I hope the organizers will remove it. Personally, I hope it does not have that property.
\n\nPlease clarify @cristiancanton
", "rawMarkdown": "Long time ago I realized that by a simple frame difference rmse clustering, one can detect pairs of (true, fake) videos that are from the same original video. By that way, we can simply correct predictions of each pair. I guess this is not what the organizers want, because it's tricky.\n\nAs a result, does the private test set has that property? If yes, then the competition solution will be tricky, and I hope the organizers will remove it. Personally, I hope it does not have that property. \n\nPlease clarify @cristiancanton ", "votes": 2 }, { "id": 791782, "postDate": "2020-03-30T16:49:45.833Z", "content": "There is a deleted comment, can somebody tell me what that was about?
", "rawMarkdown": "There is a deleted comment, can somebody tell me what that was about?", "replies": [ { "id": 791788, "postDate": "2020-03-30T16:53:49.180Z", "content": "it was off-topic. About overfitting
", "rawMarkdown": "it was off-topic. About overfitting ", "votes": 1 }, { "id": 791817, "postDate": "2020-03-30T17:30:00.110Z", "content": "I deleted it because I didn't express my idea clearly.
", "rawMarkdown": "I deleted it because I didn't express my idea clearly. ", "votes": 1 } ] }, { "id": 791767, "postDate": "2020-03-30T16:37:32.433Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 791815, "author_name": "nosound", "author_url": "", "post_date": "2020-03-30T17:27:46.220000", "content": "Not a 100% proof, but I checked by LB probing whether there are identical audio hashes in the public test, and there are none. All audio hashes are different. It (almost) proofs that the public test dataset does not contain such pairs.
", "votes": 3, "replies": [] }, { "id": 791776, "author_name": "Igor Vikhrov", "author_url": "", "post_date": "2020-03-30T16:43:51.277000", "content": "nice idea! But obviously it's cheat and does not worth money because this trick does not decide problem of detecting fake faces in real life.
", "votes": 1, "replies": [] }, { "id": 791782, "author_name": "Kha Vo", "author_url": "", "post_date": "2020-03-30T16:49:45.833000", "content": "There is a deleted comment, can somebody tell me what that was about?
", "votes": 0, "replies": [ { "id": 791788, "author_name": "Igor Vikhrov", "author_url": "", "post_date": "2020-03-30T16:53:49.180000", "content": "it was off-topic. About overfitting
", "votes": 1, "replies": [] }, { "id": 791817, "author_name": "yuanzhe zhou", "author_url": "", "post_date": "2020-03-30T17:30:00.110000", "content": "I deleted it because I didn't express my idea clearly.
", "votes": 1, "replies": [] } ] }, { "id": 791767, "author_name": "", "author_url": "", "post_date": "2020-03-30T16:37:32.433000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "791815": "Not a 100% proof, but I checked by LB probing whether there are identical audio hashes in the public test, and there are none. All audio hashes are different. It (almost) proofs that the public test dataset does not contain such pairs.", "791776": "nice idea! But obviously it's cheat and does not worth money because this trick does not decide problem of detecting fake faces in real life. ", "791756": "Long time ago I realized that by a simple frame difference rmse clustering, one can detect pairs of (true, fake) videos that are from the same original video. By that way, we can simply correct predictions of each pair. I guess this is not what the organizers want, because it's tricky.\n\nAs a result, does the private test set has that property? If yes, then the competition solution will be tricky, and I hope the organizers will remove it. Personally, I hope it does not have that property. \n\nPlease clarify @cristiancanton ", "791782": "There is a deleted comment, can somebody tell me what that was about?", "791767": "" } }