Let's calculate.. to detect face on one frame in cv pacage takes ~.15 s , we have 300 frames in video, so it is 45s for one video. we have 400 test_set video 400*45s=5h. And you cannot save these results to inference next time, because we have not privet test set. What do you think about that? :)
", "messages": [ { "id": "694356", "postDate": "12/13/2019 13:11:20", "content": "Let's calculate.. to detect face on one frame in cv pacage takes ~.15 s , we have 300 frames in video, so it is 45s for one video. we have 400 test_set video 400*45s=5h. And you cannot save these results to inference next time, because we have not privet test set. What do you think about that? :)
", "rawMarkdown": "Let's calculate.. to detect face on one frame in cv pacage takes ~.15 s , we have 300 frames in video, so it is 45s for one video. we have 400 test_set video 400*45s=5h. And you cannot save these results to inference next time, because we have not privet test set. What do you think about that? :)", "votes": null }, { "id": "694427", "postDate": "12/13/2019 15:21:17", "content": "You don't need to infere the whole video in order to say if it is fake or not. I'm thinking on something like that:
\n\n~~~\nfor frame in video:\n - Predict frame\n - Update accumulated_confidence with prediction\n If accumulated_confidence > threshold:\n break
\n\nreturn accumulated_confidence > threshold\n~~~
", "rawMarkdown": "You don't need to infere the whole video in order to say if it is fake or not. I'm thinking on something like that:\n\n~~~\nfor frame in video:\n - Predict frame\n - Update accumulated_confidence with prediction\n If accumulated_confidence > threshold:\n break\n\nreturn accumulated_confidence > threshold\n~~~", "votes": null }, { "id": "694474", "postDate": "12/13/2019 17:02:59", "content": "predict with confidence intervals? how do you evaluate confidence ?
", "rawMarkdown": "predict with confidence intervals? how do you evaluate confidence ?", "votes": null }, { "id": "694509", "postDate": "12/13/2019 18:10:14", "content": "It is only an idea that crossed my mind. For example you keep a counter for number_of_frames_predicted and other with the predictions_total.
Firstly you predict N frames on a batch, if predictions_total/number_of_frames_predicted > upper_threshold then the whole video is fake, else if predictions_total/number_of_frames_predicted < lower_threshold then the whole video is real otherwise you predict and extra frame. Keep iterating untill you meet any conditions or run out of frames, in that case it would be real.
Oh btw, the 0.15s \"cv\" stuff you're talking about is not going to work anyway... the only one that works decently on my side is about 0.5s on a gtx1080ti.... this IS an issue we have to address. Sure.
", "rawMarkdown": "Oh btw, the 0.15s \"cv\" stuff you're talking about is not going to work anyway... the only one that works decently on my side is about 0.5s on a gtx1080ti.... this IS an issue we have to address. Sure.", "votes": null }, { "id": "694626", "postDate": "12/13/2019 21:23:12", "content": "you are right, that 0.15s was on my cpu
", "rawMarkdown": "you are right, that 0.15s was on my cpu", "votes": null }, { "id": "694628", "postDate": "12/13/2019 21:27:25", "content": "btw.. my calculation took place ~3h on public validation set, on private test set timed out after >9 h. that's mean private test set is >3x bigger... and I thought that 9h limitation is for public validation set, no I am don't know.
", "rawMarkdown": "btw.. my calculation took place ~3h on public validation set, on private test set timed out after >9 h. that's mean private test set is >3x bigger... and I thought that 9h limitation is for public validation set, no I am don't know.", "votes": null }, { "id": "695213", "postDate": "12/14/2019 19:15:06", "content": "That's a great idea, @cayala ! Did you implement it?
", "rawMarkdown": "That's a great idea, @cayala ! Did you implement it?", "votes": null }, { "id": "695558", "postDate": "12/15/2019 10:42:32", "content": "I haven't started with this competition yet, I'm planning to do it this week. I'll let you know if it works.
", "rawMarkdown": "I haven't started with this competition yet, I'm planning to do it this week. I'll let you know if it works.", "votes": null } ], "comments": [ { "id": 694427, "author_name": "cayala", "author_url": "", "post_date": "12/13/2019 15:21:17", "content": "You don't need to infere the whole video in order to say if it is fake or not. I'm thinking on something like that:
\n\n~~~\nfor frame in video:\n - Predict frame\n - Update accumulated_confidence with prediction\n If accumulated_confidence > threshold:\n break
\n\nreturn accumulated_confidence > threshold\n~~~
", "votes": null, "replies": [ { "id": 694474, "author_name": "jonasmatuzas", "author_url": "", "post_date": "12/13/2019 17:02:59", "content": "predict with confidence intervals? how do you evaluate confidence ?
", "votes": null, "replies": [] }, { "id": 694509, "author_name": "cayala", "author_url": "", "post_date": "12/13/2019 18:10:14", "content": "It is only an idea that crossed my mind. For example you keep a counter for number_of_frames_predicted and other with the predictions_total.
Firstly you predict N frames on a batch, if predictions_total/number_of_frames_predicted > upper_threshold then the whole video is fake, else if predictions_total/number_of_frames_predicted < lower_threshold then the whole video is real otherwise you predict and extra frame. Keep iterating untill you meet any conditions or run out of frames, in that case it would be real.
That's a great idea, @cayala ! Did you implement it?
", "votes": null, "replies": [] }, { "id": 695558, "author_name": "cayala", "author_url": "", "post_date": "12/15/2019 10:42:32", "content": "I haven't started with this competition yet, I'm planning to do it this week. I'll let you know if it works.
", "votes": null, "replies": [] } ] }, { "id": 694598, "author_name": "simoninparis", "author_url": "", "post_date": "12/13/2019 20:28:10", "content": "Oh btw, the 0.15s \"cv\" stuff you're talking about is not going to work anyway... the only one that works decently on my side is about 0.5s on a gtx1080ti.... this IS an issue we have to address. Sure.
", "votes": null, "replies": [ { "id": 694626, "author_name": "jonasmatuzas", "author_url": "", "post_date": "12/13/2019 21:23:12", "content": "you are right, that 0.15s was on my cpu
", "votes": null, "replies": [] } ] }, { "id": 694628, "author_name": "jonasmatuzas", "author_url": "", "post_date": "12/13/2019 21:27:25", "content": "btw.. my calculation took place ~3h on public validation set, on private test set timed out after >9 h. that's mean private test set is >3x bigger... and I thought that 9h limitation is for public validation set, no I am don't know.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "694356": "Let's calculate.. to detect face on one frame in cv pacage takes ~.15 s , we have 300 frames in video, so it is 45s for one video. we have 400 test_set video 400*45s=5h. And you cannot save these results to inference next time, because we have not privet test set. What do you think about that? :)", "694427": "You don't need to infere the whole video in order to say if it is fake or not. I'm thinking on something like that:\n\n~~~\nfor frame in video:\n - Predict frame\n - Update accumulated_confidence with prediction\n If accumulated_confidence > threshold:\n break\n\nreturn accumulated_confidence > threshold\n~~~", "694474": "predict with confidence intervals? how do you evaluate confidence ?", "694509": "It is only an idea that crossed my mind. For example you keep a counter for `number_of_frames_predicted` and other with the `predictions_total`. \n\nFirstly you predict N frames on a batch, `if predictions_total/number_of_frames_predicted > upper_threshold` then the whole video is fake, `else if predictions_total/number_of_frames_predicted < lower_threshold` then the whole video is real otherwise you predict and extra frame. Keep iterating untill you meet any conditions or run out of frames, in that case it would be real.", "694598": "Oh btw, the 0.15s \"cv\" stuff you're talking about is not going to work anyway... the only one that works decently on my side is about 0.5s on a gtx1080ti.... this IS an issue we have to address. Sure.", "694626": "you are right, that 0.15s was on my cpu", "694628": "btw.. my calculation took place ~3h on public validation set, on private test set timed out after >9 h. that's mean private test set is >3x bigger... and I thought that 9h limitation is for public validation set, no I am don't know.", "695213": "That's a great idea, @cayala ! Did you implement it?", "695558": "I haven't started with this competition yet, I'm planning to do it this week. I'll let you know if it works." }, "source": "meta" }