I think it would be nice if we can share our input format; in my case, it's image as I just finished separating frames from videos. How about you?
", "messages": [ { "id": "706994", "postDate": "12/31/2019 06:11:52", "content": "I think it would be nice if we can share our input format; in my case, it's image as I just finished separating frames from videos. How about you?
", "rawMarkdown": "I think it would be nice if we can share our input format; in my case, it's image as I just finished separating frames from videos. How about you?", "votes": null }, { "id": "707143", "postDate": "12/31/2019 10:18:56", "content": "Have you try to detect face and use only face to put into CNN? Maybe audio in the video could have much information to try.
", "rawMarkdown": "Have you try to detect face and use only face to put into CNN? Maybe audio in the video could have much information to try.", "votes": null }, { "id": "707151", "postDate": "12/31/2019 10:32:19", "content": "Until now, my concept was: video -> frame(s) -> face detection -> CNN real/fake classifier. But the prototype does not look very promising. Currently, all the models I tried out converge to a \"0.5\" probability, hinting that they are unable to distinguish between real or fake. I'm also unsure whether increasing/improving the training data and/or tweaking the CNN models would lead to better results. ...it is a challenge indeed.
", "rawMarkdown": "Until now, my concept was: video -> frame(s) -> face detection -> CNN real/fake classifier. But the prototype does not look very promising. Currently, all the models I tried out converge to a \"0.5\" probability, hinting that they are unable to distinguish between real or fake. I'm also unsure whether increasing/improving the training data and/or tweaking the CNN models would lead to better results. ...it is a challenge indeed.", "votes": null }, { "id": "707281", "postDate": "12/31/2019 14:54:50", "content": "One of the reasons I am interested in this competition is that it uses videos. It seems \"obvious\" that the temporal information in the video is useful for this kind of task.
\n\nFor example, in a number of the fake videos the faces are not very consistent between frames. So one approach could be to detect face landmarks (in 3D) for every frame, and compare them between frames. If the 3D model changes too much from one frame to the next, the video is fake. Or something like that. 😄
", "rawMarkdown": "One of the reasons I am interested in this competition is that it uses videos. It seems \"obvious\" that the temporal information in the video is useful for this kind of task. \n\nFor example, in a number of the fake videos the faces are not very consistent between frames. So one approach could be to detect face landmarks (in 3D) for every frame, and compare them between frames. If the 3D model changes too much from one frame to the next, the video is fake. Or something like that. 😄", "votes": null }, { "id": "709745", "postDate": "01/03/2020 22:18:58", "content": "I'm doing same thing as @dagnelies. But like @humananalog suggested, I believe the sequential model would be useful somehow
", "rawMarkdown": "I'm doing same thing as @dagnelies. But like @humananalog suggested, I believe the sequential model would be useful somehow", "votes": null } ], "comments": [ { "id": 707143, "author_name": "phamquocbao", "author_url": "", "post_date": "12/31/2019 10:18:56", "content": "Have you try to detect face and use only face to put into CNN? Maybe audio in the video could have much information to try.
", "votes": null, "replies": [] }, { "id": 707151, "author_name": "dagnelies", "author_url": "", "post_date": "12/31/2019 10:32:19", "content": "Until now, my concept was: video -> frame(s) -> face detection -> CNN real/fake classifier. But the prototype does not look very promising. Currently, all the models I tried out converge to a \"0.5\" probability, hinting that they are unable to distinguish between real or fake. I'm also unsure whether increasing/improving the training data and/or tweaking the CNN models would lead to better results. ...it is a challenge indeed.
", "votes": null, "replies": [] }, { "id": 707281, "author_name": "humananalog", "author_url": "", "post_date": "12/31/2019 14:54:50", "content": "One of the reasons I am interested in this competition is that it uses videos. It seems \"obvious\" that the temporal information in the video is useful for this kind of task.
\n\nFor example, in a number of the fake videos the faces are not very consistent between frames. So one approach could be to detect face landmarks (in 3D) for every frame, and compare them between frames. If the 3D model changes too much from one frame to the next, the video is fake. Or something like that. 😄
", "votes": null, "replies": [] }, { "id": 709745, "author_name": "", "author_url": "", "post_date": "01/03/2020 22:18:58", "content": "I'm doing same thing as @dagnelies. But like @humananalog suggested, I believe the sequential model would be useful somehow
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "706994": "I think it would be nice if we can share our input format; in my case, it's image as I just finished separating frames from videos. How about you?", "707143": "Have you try to detect face and use only face to put into CNN? Maybe audio in the video could have much information to try.", "707151": "Until now, my concept was: video -> frame(s) -> face detection -> CNN real/fake classifier. But the prototype does not look very promising. Currently, all the models I tried out converge to a \"0.5\" probability, hinting that they are unable to distinguish between real or fake. I'm also unsure whether increasing/improving the training data and/or tweaking the CNN models would lead to better results. ...it is a challenge indeed.", "707281": "One of the reasons I am interested in this competition is that it uses videos. It seems \"obvious\" that the temporal information in the video is useful for this kind of task. \n\nFor example, in a number of the fake videos the faces are not very consistent between frames. So one approach could be to detect face landmarks (in 3D) for every frame, and compare them between frames. If the 3D model changes too much from one frame to the next, the video is fake. Or something like that. 😄", "709745": "I'm doing same thing as @dagnelies. But like @humananalog suggested, I believe the sequential model would be useful somehow" }, "source": "meta" }