From the Machine Learning point of view, any process that generate train data and test data by 2 different ways will damage test predictions.
\n\nI see some discussion saying that we can use face boxes detected in real videos to translate into corresponding fake videos. By this way training fake boxes are good. But in test time we cannot perform this process. Test faces will be generated independently for both real and fake. So for fake test videos, the process is not the same.
\n\nShould we embrace the noise in training data by just normally performing face detector independently for real and fake training videos?
\n\nIMHO, we should do that instead of sophisticating the process.
", "messages": [ { "id": "765904", "postDate": "03/07/2020 10:33:10", "content": "From the Machine Learning point of view, any process that generate train data and test data by 2 different ways will damage test predictions.
\n\nI see some discussion saying that we can use face boxes detected in real videos to translate into corresponding fake videos. By this way training fake boxes are good. But in test time we cannot perform this process. Test faces will be generated independently for both real and fake. So for fake test videos, the process is not the same.
\n\nShould we embrace the noise in training data by just normally performing face detector independently for real and fake training videos?
\n\nIMHO, we should do that instead of sophisticating the process.
", "rawMarkdown": "From the Machine Learning point of view, any process that generate train data and test data by 2 different ways will damage test predictions.\n\nI see some discussion saying that we can use face boxes detected in real videos to translate into corresponding fake videos. By this way training fake boxes are good. But in test time we cannot perform this process. Test faces will be generated independently for both real and fake. So for fake test videos, the process is not the same.\n\nShould we embrace the noise in training data by just normally performing face detector independently for real and fake training videos? \n\nIMHO, we should do that instead of sophisticating the process.", "votes": null }, { "id": "767045", "postDate": "03/09/2020 04:43:22", "content": "I think there are multiple factors there. I think for some people that are doing that in the data prep phase for the sake of reducing compute cost, if you only face detect one video instead of multiple it is much less work. Others are doing that to do a comparison to make sure that the facial regions are actually being changed between the real and fake in the region we expect.
\n\nI understand your concern, but I don't think that it will truly make a material difference. I have not noticed any difference of facial recognition between real and fake videos. It seems to find roughly similar points for both. If it does make a difference I would wonder exactly how much error might be introduced from alternate facial recognition. Seems it would be relatively minor.
\n\nI have not done that method, early on I did a single run through all videos, but in hindsight I think only running it on the real videos was probably viable.
", "rawMarkdown": "I think there are multiple factors there. I think for some people that are doing that in the data prep phase for the sake of reducing compute cost, if you only face detect one video instead of multiple it is much less work. Others are doing that to do a comparison to make sure that the facial regions are actually being changed between the real and fake in the region we expect. \n\nI understand your concern, but I don't think that it will truly make a material difference. I have not noticed any difference of facial recognition between real and fake videos. It seems to find roughly similar points for both. If it does make a difference I would wonder exactly how much error might be introduced from alternate facial recognition. Seems it would be relatively minor. \n\nI have not done that method, early on I did a single run through all videos, but in hindsight I think only running it on the real videos was probably viable.", "votes": null }, { "id": "767153", "postDate": "03/09/2020 08:31:08", "content": "During 1 past competition, I experienced this issue before. A perfect test feature set is not as good in test predictions as the test features retrieved by generating with the same process as train features. The model can learn the noise and perform better. But I may be wrong.
", "rawMarkdown": "During 1 past competition, I experienced this issue before. A perfect test feature set is not as good in test predictions as the test features retrieved by generating with the same process as train features. The model can learn the noise and perform better. But I may be wrong.", "votes": null }, { "id": "767157", "postDate": "03/09/2020 08:38:52", "content": "It's definitely possible that a fake messes with the face so strongly that it makes it so the facial recognition doesn't accurately finds it and finds some other random region instead and if you dont have this in the training process then it wont know how to respond to that alternate input, but I have not seen that very commonly and with how most people are doing many frames of prediction I think those would likely be smoothed out anyway
", "rawMarkdown": "It's definitely possible that a fake messes with the face so strongly that it makes it so the facial recognition doesn't accurately finds it and finds some other random region instead and if you dont have this in the training process then it wont know how to respond to that alternate input, but I have not seen that very commonly and with how most people are doing many frames of prediction I think those would likely be smoothed out anyway", "votes": null }, { "id": "767589", "postDate": "03/09/2020 21:00:45", "content": "@khahuras, I was one of them (or the only one) who said that \"you can transfer a box from real to fake video to get a better bounding box\", well, another thing you can do to solve your problem is just simply add a random number, which will slightly move the bounding box, so the face is still in there and it also have that effect of classified by face detector. EZ
", "rawMarkdown": "khahuras, I was one of them (or the only one) who said that \"you can transfer a box from real to fake video to get a better bounding box\", well, another thing you can do to solve your problem is just simply add a random number, which will slightly move the bounding box, so the face is still in there and it also have that effect of classified by face detector. EZ", "votes": null } ], "comments": [ { "id": 767045, "author_name": "ryches", "author_url": "", "post_date": "03/09/2020 04:43:22", "content": "I think there are multiple factors there. I think for some people that are doing that in the data prep phase for the sake of reducing compute cost, if you only face detect one video instead of multiple it is much less work. Others are doing that to do a comparison to make sure that the facial regions are actually being changed between the real and fake in the region we expect.
\n\nI understand your concern, but I don't think that it will truly make a material difference. I have not noticed any difference of facial recognition between real and fake videos. It seems to find roughly similar points for both. If it does make a difference I would wonder exactly how much error might be introduced from alternate facial recognition. Seems it would be relatively minor.
\n\nI have not done that method, early on I did a single run through all videos, but in hindsight I think only running it on the real videos was probably viable.
", "votes": null, "replies": [ { "id": 767153, "author_name": "khahuras", "author_url": "", "post_date": "03/09/2020 08:31:08", "content": "During 1 past competition, I experienced this issue before. A perfect test feature set is not as good in test predictions as the test features retrieved by generating with the same process as train features. The model can learn the noise and perform better. But I may be wrong.
", "votes": null, "replies": [] }, { "id": 767157, "author_name": "ryches", "author_url": "", "post_date": "03/09/2020 08:38:52", "content": "It's definitely possible that a fake messes with the face so strongly that it makes it so the facial recognition doesn't accurately finds it and finds some other random region instead and if you dont have this in the training process then it wont know how to respond to that alternate input, but I have not seen that very commonly and with how most people are doing many frames of prediction I think those would likely be smoothed out anyway
", "votes": null, "replies": [] } ] }, { "id": 767589, "author_name": "harshitsheoran", "author_url": "", "post_date": "03/09/2020 21:00:45", "content": "@khahuras, I was one of them (or the only one) who said that \"you can transfer a box from real to fake video to get a better bounding box\", well, another thing you can do to solve your problem is just simply add a random number, which will slightly move the bounding box, so the face is still in there and it also have that effect of classified by face detector. EZ
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "765904": "From the Machine Learning point of view, any process that generate train data and test data by 2 different ways will damage test predictions.\n\nI see some discussion saying that we can use face boxes detected in real videos to translate into corresponding fake videos. By this way training fake boxes are good. But in test time we cannot perform this process. Test faces will be generated independently for both real and fake. So for fake test videos, the process is not the same.\n\nShould we embrace the noise in training data by just normally performing face detector independently for real and fake training videos? \n\nIMHO, we should do that instead of sophisticating the process.", "767045": "I think there are multiple factors there. I think for some people that are doing that in the data prep phase for the sake of reducing compute cost, if you only face detect one video instead of multiple it is much less work. Others are doing that to do a comparison to make sure that the facial regions are actually being changed between the real and fake in the region we expect. \n\nI understand your concern, but I don't think that it will truly make a material difference. I have not noticed any difference of facial recognition between real and fake videos. It seems to find roughly similar points for both. If it does make a difference I would wonder exactly how much error might be introduced from alternate facial recognition. Seems it would be relatively minor. \n\nI have not done that method, early on I did a single run through all videos, but in hindsight I think only running it on the real videos was probably viable.", "767153": "During 1 past competition, I experienced this issue before. A perfect test feature set is not as good in test predictions as the test features retrieved by generating with the same process as train features. The model can learn the noise and perform better. But I may be wrong.", "767157": "It's definitely possible that a fake messes with the face so strongly that it makes it so the facial recognition doesn't accurately finds it and finds some other random region instead and if you dont have this in the training process then it wont know how to respond to that alternate input, but I have not seen that very commonly and with how most people are doing many frames of prediction I think those would likely be smoothed out anyway", "767589": "khahuras, I was one of them (or the only one) who said that \"you can transfer a box from real to fake video to get a better bounding box\", well, another thing you can do to solve your problem is just simply add a random number, which will slightly move the bounding box, so the face is still in there and it also have that effect of classified by face detector. EZ" }, "source": "meta" }