![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F4215732%2F73af66eebaadc2f5439523b43bdf6d74%2Fzoxhjusmnw_0.jpg?generation=1583598160359781&alt=media\")
I am curious if anyone has been able to extract all the faces from all the videos (around 119146)?\nI have been running an extraction script for around 2 days and got so far around 4000 videos processed. \nI guess I should stop at some point and train with what I got.\nAny thoughts on that?
", "messages": [ { "id": "764173", "postDate": "03/05/2020 07:44:34", "content": "I am curious if anyone has been able to extract all the faces from all the videos (around 119146)?\nI have been running an extraction script for around 2 days and got so far around 4000 videos processed. \nI guess I should stop at some point and train with what I got.\nAny thoughts on that?
", "rawMarkdown": "I am curious if anyone has been able to extract all the faces from all the videos (around 119146)?\nI have been running an extraction script for around 2 days and got so far around 4000 videos processed. \nI guess I should stop at some point and train with what I got.\nAny thoughts on that?", "votes": null }, { "id": "764214", "postDate": "03/05/2020 08:21:49", "content": "You should try strides like in mtcnn kernel. Try retina face with striding. And in time you should try to train model on datasets stated in \"other useful datasets\" named discussion. They're well created. Use other method for reading video, there's a kernel named something like \"decorder\" which seems to be pretty fast. By using all these you should speed it up and get your dataset ready within 5 days.
", "rawMarkdown": "You should try strides like in mtcnn kernel. Try retina face with striding. And in time you should try to train model on datasets stated in \"other useful datasets\" named discussion. They're well created. Use other method for reading video, there's a kernel named something like \"decorder\" which seems to be pretty fast. By using all these you should speed it up and get your dataset ready within 5 days.", "votes": null }, { "id": "764233", "postDate": "03/05/2020 08:44:35", "content": "Please dont toture your pc and yourself doing all faces, it does not make a difference after 10 frames from each, but if you wanna go further you can choose 30 - 120 frames depending on the video, as done by james howard (currently 5th on the leaderboard).
", "rawMarkdown": "Please dont toture your pc and yourself doing all faces, it does not make a difference after 10 frames from each, but if you wanna go further you can choose 30 - 120 frames depending on the video, as done by james howard (currently 5th on the leaderboard).", "votes": null }, { "id": "764295", "postDate": "03/05/2020 10:02:26", "content": "Are you running a face detector on all 120K videos? because you can just run it on the real videos and use the same bounding boxes to crop the fake faces.
", "rawMarkdown": "Are you running a face detector on all 120K videos? because you can just run it on the real videos and use the same bounding boxes to crop the fake faces.", "votes": null }, { "id": "764399", "postDate": "03/05/2020 12:22:15", "content": "Thanks for the suggestion. Any method you suggest for selecting these 30-120 frames ?
", "rawMarkdown": "Thanks for the suggestion. Any method you suggest for selecting these 30-120 frames ?", "votes": null }, { "id": "764400", "postDate": "03/05/2020 12:22:42", "content": "That's a great idea, will look into it! Thanks for the hint.
", "rawMarkdown": "That's a great idea, will look into it! Thanks for the hint.", "votes": null }, { "id": "764654", "postDate": "03/05/2020 17:42:51", "content": "Even I dont know what he is doing exactly, but I guess maybe he is taking more frames from the real one and less from the fake one, you can get more with this line \"one face only comes one time\" which he wrote, definately not very easy to interpret. Think about it I guess you will get how he might be doing it.
", "rawMarkdown": "Even I dont know what he is doing exactly, but I guess maybe he is taking more frames from the real one and less from the fake one, you can get more with this line \"one face only comes one time\" which he wrote, definately not very easy to interpret. Think about it I guess you will get how he might be doing it.", "votes": null }, { "id": "765175", "postDate": "03/06/2020 10:22:16", "content": "did you all the faces from all the videos. ho much time doe it took for all and what's your RAM size and hdd?
", "rawMarkdown": "did you all the faces from all the videos. ho much time doe it took for all and what's your RAM size and hdd?", "votes": null }, { "id": "765187", "postDate": "03/06/2020 10:47:05", "content": "I've stopped at around 5000 videos and have extracted all the frames (around 300).\nI am now working on extracting less frames and being mort \"smart\" about how I do it. ;)
\nMy setup is 64GB of RAM, 8 CPU cores, and 2TB of HDD. The RAM usage is quite low. \nHope this helps!
Extract only from real videos, extract every 5/10 frames then interpolate results. Whole-time for my system (64 RAM, 1080TI, Threadripper 2920) - 10 hours.
", "rawMarkdown": "Extract only from real videos, extract every 5/10 frames then interpolate results. Whole-time for my system (64 RAM, 1080TI, Threadripper 2920) - 10 hours.", "votes": null }, { "id": "766024", "postDate": "03/07/2020 15:13:24", "content": "[UPDATE] Here is how the new face detection pipeline will look like:
\n\nI will post the code once it is ready in the following thread: https://www.kaggle.com/c/deepfake-detection-challenge/discussion/133225
", "rawMarkdown": "[UPDATE] Here is how the new face detection pipeline will look like: \n\n- Bounding box extraction for real videos for 20 to 25 frames (instead of 300)\n- Saving these as a 4 numbers in a Parquet file\n- Cropping the faces from real and fake videos\n\nI will post the code once it is ready in the following thread: https://www.kaggle.com/c/deepfake-detection-challenge/discussion/133225", "votes": null }, { "id": "766025", "postDate": "03/07/2020 15:13:52", "content": "Awesome, thanks for sharing the details of your results!
", "rawMarkdown": "Awesome, thanks for sharing the details of your results!", "votes": null }, { "id": "766078", "postDate": "03/07/2020 16:25:44", "content": "Just to keep in mind fake faces are not always at face locations(there are several different examples like that):\n
Edit: watch video zoxhjusmnw.mp4 from part 36 the face is on the clouds then goes to trousers
", "rawMarkdown": "Just to keep in mind fake faces are not always at face locations(there are several different examples like that):\n\n\nEdit: watch video zoxhjusmnw.mp4 from part 36 the face is on the clouds then goes to trousers", "votes": null }, { "id": "766102", "postDate": "03/07/2020 16:59:30", "content": "Great point indeed! What do you recommend for such samples? Or maybe you prefer to reveal your tricks only after the competition ends? ;)
", "rawMarkdown": "Great point indeed! What do you recommend for such samples? Or maybe you prefer to reveal your tricks only after the competition ends? ;)", "votes": null } ], "comments": [ { "id": 764214, "author_name": "ankitsainiankit", "author_url": "", "post_date": "03/05/2020 08:21:49", "content": "You should try strides like in mtcnn kernel. Try retina face with striding. And in time you should try to train model on datasets stated in \"other useful datasets\" named discussion. They're well created. Use other method for reading video, there's a kernel named something like \"decorder\" which seems to be pretty fast. By using all these you should speed it up and get your dataset ready within 5 days.
", "votes": null, "replies": [] }, { "id": 764233, "author_name": "harshitsheoran", "author_url": "", "post_date": "03/05/2020 08:44:35", "content": "Please dont toture your pc and yourself doing all faces, it does not make a difference after 10 frames from each, but if you wanna go further you can choose 30 - 120 frames depending on the video, as done by james howard (currently 5th on the leaderboard).
", "votes": null, "replies": [ { "id": 764399, "author_name": "yassinealouini", "author_url": "", "post_date": "03/05/2020 12:22:15", "content": "Thanks for the suggestion. Any method you suggest for selecting these 30-120 frames ?
", "votes": null, "replies": [] }, { "id": 764654, "author_name": "harshitsheoran", "author_url": "", "post_date": "03/05/2020 17:42:51", "content": "Even I dont know what he is doing exactly, but I guess maybe he is taking more frames from the real one and less from the fake one, you can get more with this line \"one face only comes one time\" which he wrote, definately not very easy to interpret. Think about it I guess you will get how he might be doing it.
", "votes": null, "replies": [] } ] }, { "id": 764295, "author_name": "ma7moud", "author_url": "", "post_date": "03/05/2020 10:02:26", "content": "Are you running a face detector on all 120K videos? because you can just run it on the real videos and use the same bounding boxes to crop the fake faces.
", "votes": null, "replies": [ { "id": 764400, "author_name": "yassinealouini", "author_url": "", "post_date": "03/05/2020 12:22:42", "content": "That's a great idea, will look into it! Thanks for the hint.
", "votes": null, "replies": [] }, { "id": 766078, "author_name": "emrebayram", "author_url": "", "post_date": "03/07/2020 16:25:44", "content": "Just to keep in mind fake faces are not always at face locations(there are several different examples like that):\n
Edit: watch video zoxhjusmnw.mp4 from part 36 the face is on the clouds then goes to trousers
", "votes": null, "replies": [] }, { "id": 766102, "author_name": "yassinealouini", "author_url": "", "post_date": "03/07/2020 16:59:30", "content": "Great point indeed! What do you recommend for such samples? Or maybe you prefer to reveal your tricks only after the competition ends? ;)
", "votes": null, "replies": [] } ] }, { "id": 765175, "author_name": "anushakarthik1991", "author_url": "", "post_date": "03/06/2020 10:22:16", "content": "did you all the faces from all the videos. ho much time doe it took for all and what's your RAM size and hdd?
", "votes": null, "replies": [ { "id": 765187, "author_name": "yassinealouini", "author_url": "", "post_date": "03/06/2020 10:47:05", "content": "I've stopped at around 5000 videos and have extracted all the frames (around 300).\nI am now working on extracting less frames and being mort \"smart\" about how I do it. ;)
\nMy setup is 64GB of RAM, 8 CPU cores, and 2TB of HDD. The RAM usage is quite low. \nHope this helps!
Extract only from real videos, extract every 5/10 frames then interpolate results. Whole-time for my system (64 RAM, 1080TI, Threadripper 2920) - 10 hours.
", "votes": null, "replies": [ { "id": 766025, "author_name": "yassinealouini", "author_url": "", "post_date": "03/07/2020 15:13:52", "content": "Awesome, thanks for sharing the details of your results!
", "votes": null, "replies": [] } ] }, { "id": 766024, "author_name": "yassinealouini", "author_url": "", "post_date": "03/07/2020 15:13:24", "content": "[UPDATE] Here is how the new face detection pipeline will look like:
\n\nI will post the code once it is ready in the following thread: https://www.kaggle.com/c/deepfake-detection-challenge/discussion/133225
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "764173": "I am curious if anyone has been able to extract all the faces from all the videos (around 119146)?\nI have been running an extraction script for around 2 days and got so far around 4000 videos processed. \nI guess I should stop at some point and train with what I got.\nAny thoughts on that?", "764214": "You should try strides like in mtcnn kernel. Try retina face with striding. And in time you should try to train model on datasets stated in \"other useful datasets\" named discussion. They're well created. Use other method for reading video, there's a kernel named something like \"decorder\" which seems to be pretty fast. By using all these you should speed it up and get your dataset ready within 5 days.", "764233": "Please dont toture your pc and yourself doing all faces, it does not make a difference after 10 frames from each, but if you wanna go further you can choose 30 - 120 frames depending on the video, as done by james howard (currently 5th on the leaderboard).", "764295": "Are you running a face detector on all 120K videos? because you can just run it on the real videos and use the same bounding boxes to crop the fake faces.", "764399": "Thanks for the suggestion. Any method you suggest for selecting these 30-120 frames ?", "764400": "That's a great idea, will look into it! Thanks for the hint.", "764654": "Even I dont know what he is doing exactly, but I guess maybe he is taking more frames from the real one and less from the fake one, you can get more with this line \"one face only comes one time\" which he wrote, definately not very easy to interpret. Think about it I guess you will get how he might be doing it.", "765175": "did you all the faces from all the videos. ho much time doe it took for all and what's your RAM size and hdd?", "765187": "I've stopped at around 5000 videos and have extracted all the frames (around 300).\nI am now working on extracting less frames and being mort \"smart\" about how I do it. ;) \nMy setup is 64GB of RAM, 8 CPU cores, and 2TB of HDD. The RAM usage is quite low. \nHope this helps!", "765984": "Extract only from real videos, extract every 5/10 frames then interpolate results. Whole-time for my system (64 RAM, 1080TI, Threadripper 2920) - 10 hours.", "766024": "[UPDATE] Here is how the new face detection pipeline will look like: \n\n- Bounding box extraction for real videos for 20 to 25 frames (instead of 300)\n- Saving these as a 4 numbers in a Parquet file\n- Cropping the faces from real and fake videos\n\nI will post the code once it is ready in the following thread: https://www.kaggle.com/c/deepfake-detection-challenge/discussion/133225", "766025": "Awesome, thanks for sharing the details of your results!", "766078": "Just to keep in mind fake faces are not always at face locations(there are several different examples like that):\n\n\nEdit: watch video zoxhjusmnw.mp4 from part 36 the face is on the clouds then goes to trousers", "766102": "Great point indeed! What do you recommend for such samples? Or maybe you prefer to reveal your tricks only after the competition ends? ;)" }, "source": "meta" }