![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F172860%2F387c9a2d50528c4d1dc237102a7bed50%2Fvisualization.png?generation=1581854807111044&alt=media\")
Based on the metadata dataset created by @zaharch and some additional digging, here are some insights:
\n\ntrain, test, and train_sample) => {'test': 3, 'train': 50, 'train_sample': 1}
Some folders have larger mean video heights (again, check the graph below)\n.png?generation=1581855001285518&alt=media)
Some folders have larger mean video widths (again, check the graph below)
.png?generation=1581855101736718&alt=media)
Mean and std of video's number of frames are similar across the different labels: \n\n{'mean': {'FAKE': 297.3371569725578, 'REAL': 298.69938394069123},\n'std': {'FAKE': 17.675306553250675, 'REAL': 14.055158796208511}}\n
For audio's number of frames, the std is much larger for FAKE ones:
```\n{'mean': {'FAKE': 447.2821305326332, 'REAL': 466.04273548926386},\n 'std': {'FAKE': 69.65349999014951, 'REAL': 14.25591050932309}}
\n\n\n- Only one type of audio codec has been used => **aac**\n- Mean and std video durations across labels are very close:\n\n{'mean': {'FAKE': 10.012160656957871, 'REAL': 10.011291938758431},\n 'std': {'FAKE': 0.0070542637165113215, 'REAL': 0.005409740417924116}}\n```\n- Mean audio durations across labels are close but std are bigger for REAL videos:
```\n{'mean': {'FAKE': 9.999094023505874, 'REAL': 9.996256412935582},\n 'std': {'FAKE': 0.09138021378336204, 'REAL': 0.1850847423078909}}
\n\n```
\n\n.png?generation=1581856086317504&alt=media)
\n{'test': ['dfdc_train_part_2', 'dfdc_train_part_1', 'dfdc_train_part_0'],\n 'train': ['dfdc_train_part_12',\n 'dfdc_train_part_6',\n 'dfdc_train_part_21',\n 'dfdc_train_part_7',\n 'dfdc_train_part_24',\n 'dfdc_train_part_17',\n 'dfdc_train_part_39',\n 'dfdc_train_part_9',\n 'dfdc_train_part_43',\n 'dfdc_train_part_11',\n 'dfdc_train_part_23',\n 'dfdc_train_part_40',\n 'dfdc_train_part_34',\n 'dfdc_train_part_15',\n 'dfdc_train_part_2',\n 'dfdc_train_part_28',\n 'dfdc_train_part_36',\n 'dfdc_train_part_47',\n 'dfdc_train_part_32',\n 'dfdc_train_part_27',\n 'dfdc_train_part_49',\n 'dfdc_train_part_46',\n 'dfdc_train_part_35',\n 'dfdc_train_part_48',\n 'dfdc_train_part_31',\n 'dfdc_train_part_26',\n 'dfdc_train_part_42',\n 'dfdc_train_part_29',\n 'dfdc_train_part_19',\n 'dfdc_train_part_33',\n 'dfdc_train_part_37',\n 'dfdc_train_part_41',\n 'dfdc_train_part_16',\n 'dfdc_train_part_44',\n 'dfdc_train_part_1',\n 'dfdc_train_part_25',\n 'dfdc_train_part_20',\n 'dfdc_train_part_22',\n 'dfdc_train_part_10',\n 'dfdc_train_part_30',\n 'dfdc_train_part_5',\n 'dfdc_train_part_38',\n 'dfdc_train_part_4',\n 'dfdc_train_part_13',\n 'dfdc_train_part_0',\n 'dfdc_train_part_14',\n 'dfdc_train_part_45',\n 'dfdc_train_part_3',\n 'dfdc_train_part_18',\n 'dfdc_train_part_8'],\n 'train_sample': ['dfdc_train_part_28']}\n
I will add more insights the better I understand the problem + will check my code for silly bugs and correct if necessary. Let me know if you found other interesting insights.
\n\nFinally, notice that a notebook will be available at some point. ;)
", "messages": [ { "id": "747429", "postDate": "02/16/2020 12:28:49", "content": "Based on the metadata dataset created by @zaharch and some additional digging, here are some insights:
\n\ntrain, test, and train_sample) => {'test': 3, 'train': 50, 'train_sample': 1}
Some folders have larger mean video heights (again, check the graph below)\n.png?generation=1581855001285518&alt=media)
Some folders have larger mean video widths (again, check the graph below)
.png?generation=1581855101736718&alt=media)
Mean and std of video's number of frames are similar across the different labels: \n\n{'mean': {'FAKE': 297.3371569725578, 'REAL': 298.69938394069123},\n'std': {'FAKE': 17.675306553250675, 'REAL': 14.055158796208511}}\n
For audio's number of frames, the std is much larger for FAKE ones:
```\n{'mean': {'FAKE': 447.2821305326332, 'REAL': 466.04273548926386},\n 'std': {'FAKE': 69.65349999014951, 'REAL': 14.25591050932309}}
\n\n\n- Only one type of audio codec has been used => **aac**\n- Mean and std video durations across labels are very close:\n\n{'mean': {'FAKE': 10.012160656957871, 'REAL': 10.011291938758431},\n 'std': {'FAKE': 0.0070542637165113215, 'REAL': 0.005409740417924116}}\n```\n- Mean audio durations across labels are close but std are bigger for REAL videos:
```\n{'mean': {'FAKE': 9.999094023505874, 'REAL': 9.996256412935582},\n 'std': {'FAKE': 0.09138021378336204, 'REAL': 0.1850847423078909}}
\n\n```
\n\n.png?generation=1581856086317504&alt=media)
\n{'test': ['dfdc_train_part_2', 'dfdc_train_part_1', 'dfdc_train_part_0'],\n 'train': ['dfdc_train_part_12',\n 'dfdc_train_part_6',\n 'dfdc_train_part_21',\n 'dfdc_train_part_7',\n 'dfdc_train_part_24',\n 'dfdc_train_part_17',\n 'dfdc_train_part_39',\n 'dfdc_train_part_9',\n 'dfdc_train_part_43',\n 'dfdc_train_part_11',\n 'dfdc_train_part_23',\n 'dfdc_train_part_40',\n 'dfdc_train_part_34',\n 'dfdc_train_part_15',\n 'dfdc_train_part_2',\n 'dfdc_train_part_28',\n 'dfdc_train_part_36',\n 'dfdc_train_part_47',\n 'dfdc_train_part_32',\n 'dfdc_train_part_27',\n 'dfdc_train_part_49',\n 'dfdc_train_part_46',\n 'dfdc_train_part_35',\n 'dfdc_train_part_48',\n 'dfdc_train_part_31',\n 'dfdc_train_part_26',\n 'dfdc_train_part_42',\n 'dfdc_train_part_29',\n 'dfdc_train_part_19',\n 'dfdc_train_part_33',\n 'dfdc_train_part_37',\n 'dfdc_train_part_41',\n 'dfdc_train_part_16',\n 'dfdc_train_part_44',\n 'dfdc_train_part_1',\n 'dfdc_train_part_25',\n 'dfdc_train_part_20',\n 'dfdc_train_part_22',\n 'dfdc_train_part_10',\n 'dfdc_train_part_30',\n 'dfdc_train_part_5',\n 'dfdc_train_part_38',\n 'dfdc_train_part_4',\n 'dfdc_train_part_13',\n 'dfdc_train_part_0',\n 'dfdc_train_part_14',\n 'dfdc_train_part_45',\n 'dfdc_train_part_3',\n 'dfdc_train_part_18',\n 'dfdc_train_part_8'],\n 'train_sample': ['dfdc_train_part_28']}\n
I will add more insights the better I understand the problem + will check my code for silly bugs and correct if necessary. Let me know if you found other interesting insights.
\n\nFinally, notice that a notebook will be available at some point. ;)
", "rawMarkdown": "Based on the [metadata ](https://www.kaggle.com/zaharch) dataset created by @zaharch and some additional digging, here are some insights:\n\n- Here are the number of unique folders by split type (`train`, `test`, and `train_sample`) => ``` {'test': 3, 'train': 50, 'train_sample': 1} ```\n- Some folders contain more FAKE labels (check the graph below)\n\n\n\n- Some folders have larger mean video heights (again, check the graph below)\n.png?generation=1581855001285518&alt=media)\n\n- Some folders have larger mean video widths (again, check the graph below)\n\n.png?generation=1581855101736718&alt=media)\n\n- Mean and std of video's number of frames are similar across the different labels: \n```\n{'mean': {'FAKE': 297.3371569725578, 'REAL': 298.69938394069123},\n 'std': {'FAKE': 17.675306553250675, 'REAL': 14.055158796208511}}\n```\n\n- For audio's number of frames, the std is much larger for FAKE ones:\n\n```\n{'mean': {'FAKE': 447.2821305326332, 'REAL': 466.04273548926386},\n 'std': {'FAKE': 69.65349999014951, 'REAL': 14.25591050932309}}\n\n```\n- Only one type of audio codec has been used => **aac**\n- Mean and std video durations across labels are very close:\n```\n{'mean': {'FAKE': 10.012160656957871, 'REAL': 10.011291938758431},\n 'std': {'FAKE': 0.0070542637165113215, 'REAL': 0.005409740417924116}}\n```\n- Mean audio durations across labels are close but std are bigger for REAL videos:\n\n```\n{'mean': {'FAKE': 9.999094023505874, 'REAL': 9.996256412935582},\n 'std': {'FAKE': 0.09138021378336204, 'REAL': 0.1850847423078909}}\n\n```\n\n- Here is the distribution of labels across split types (I am not sure what the test split type represents since it has 50% of each?)\n\n.png?generation=1581856086317504&alt=media)\n\n- As suggested by @humananalog, here is the folder split by type (either test, train, or train_sample): \n\n```\n{'test': ['dfdc_train_part_2', 'dfdc_train_part_1', 'dfdc_train_part_0'],\n 'train': ['dfdc_train_part_12',\n 'dfdc_train_part_6',\n 'dfdc_train_part_21',\n 'dfdc_train_part_7',\n 'dfdc_train_part_24',\n 'dfdc_train_part_17',\n 'dfdc_train_part_39',\n 'dfdc_train_part_9',\n 'dfdc_train_part_43',\n 'dfdc_train_part_11',\n 'dfdc_train_part_23',\n 'dfdc_train_part_40',\n 'dfdc_train_part_34',\n 'dfdc_train_part_15',\n 'dfdc_train_part_2',\n 'dfdc_train_part_28',\n 'dfdc_train_part_36',\n 'dfdc_train_part_47',\n 'dfdc_train_part_32',\n 'dfdc_train_part_27',\n 'dfdc_train_part_49',\n 'dfdc_train_part_46',\n 'dfdc_train_part_35',\n 'dfdc_train_part_48',\n 'dfdc_train_part_31',\n 'dfdc_train_part_26',\n 'dfdc_train_part_42',\n 'dfdc_train_part_29',\n 'dfdc_train_part_19',\n 'dfdc_train_part_33',\n 'dfdc_train_part_37',\n 'dfdc_train_part_41',\n 'dfdc_train_part_16',\n 'dfdc_train_part_44',\n 'dfdc_train_part_1',\n 'dfdc_train_part_25',\n 'dfdc_train_part_20',\n 'dfdc_train_part_22',\n 'dfdc_train_part_10',\n 'dfdc_train_part_30',\n 'dfdc_train_part_5',\n 'dfdc_train_part_38',\n 'dfdc_train_part_4',\n 'dfdc_train_part_13',\n 'dfdc_train_part_0',\n 'dfdc_train_part_14',\n 'dfdc_train_part_45',\n 'dfdc_train_part_3',\n 'dfdc_train_part_18',\n 'dfdc_train_part_8'],\n 'train_sample': ['dfdc_train_part_28']}\n```\n\n\n\nI will add more insights the better I understand the problem + will check my code for silly bugs and correct if necessary. Let me know if you found other interesting insights. \n\nFinally, notice that a notebook will be available at some point. ;)", "votes": null }, { "id": "747460", "postDate": "02/16/2020 13:11:31", "content": "Just FYI, the test videos are taken from subdirs 0, 1 and 2 in the training set.
", "rawMarkdown": "Just FYI, the test videos are taken from subdirs 0, 1 and 2 in the training set.", "votes": null }, { "id": "747648", "postDate": "02/16/2020 17:35:39", "content": "thank you! @Yassine Alouini
", "rawMarkdown": "thank you! @Yassine Alouini", "votes": null }, { "id": "748346", "postDate": "02/17/2020 12:19:50", "content": "Thank you for sharing @yassinealouini ! I had a couple of questions; what do you mean by audio's number of frames and how is that different from video's number of frames?
Hey @anshulrai. To be honest, I am not sure what these quantities represent yet. \nFor now, I have quickly explored the great metadata dataset from @zaharch\nhttps://www.kaggle.com/zaharch/train-set-metadata-for-dfdc.\nI will update my exploration with more information once I have a better understanding.\nStay tuned. ;)
", "rawMarkdown": "Hey @anshulrai. To be honest, I am not sure what these quantities represent yet. \nFor now, I have quickly explored the great metadata dataset from @zaharch\nhttps://www.kaggle.com/zaharch/train-set-metadata-for-dfdc.\nI will update my exploration with more information once I have a better understanding.\nStay tuned. ;)", "votes": null }, { "id": "753546", "postDate": "02/22/2020 10:49:55", "content": "[EDIT] I was convinced that the source dataset was linked in this post but it wasn't, problem fixed now. :)
", "rawMarkdown": "[EDIT] I was convinced that the source dataset was linked in this post but it wasn't, problem fixed now. :)", "votes": null }, { "id": "753547", "postDate": "02/22/2020 10:50:07", "content": "Glad it helps!
", "rawMarkdown": "Glad it helps!", "votes": null }, { "id": "756574", "postDate": "02/25/2020 21:12:54", "content": "[UPDATE] A WIP Streamlit app. More to come in the next few days. ;)
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F172860%2Fc901b2b49fb2df9c8238354c67120b9a%2FScreenshot%20from%202020-02-25%2022-10-34.png?generation=1582665127394867&alt=media\")
Yet more progress, I really like the concept of sidebar in the streamlit API. So far, the API is clean and concise.
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F172860%2F910176e27760b13967c17dbf3063c83f%2FScreenshot%20from%202020-02-25%2023-30-57.png?generation=1582669871496852&alt=media\")
interesting, I did something similar with streamlit.\nmy app is based on prediction error, some of true/fake are impossible to tell visually
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F150338%2Fb12f2f6c1574d92d801518b440eebe61%2FQAQA15Y.png?generation=1582707100565709&alt=media\")
Awesome work! Do you mind sharing some details about the model you have there: on how many samples have you trained it, what model architecture, for how long, and so on?\nThanks in advance!
", "rawMarkdown": "Awesome work! Do you mind sharing some details about the model you have there: on how many samples have you trained it, what model architecture, for how long, and so on?\nThanks in advance!", "votes": null }, { "id": "758518", "postDate": "02/27/2020 21:16:55", "content": "Next thing to implement: adding a face detection layer => https://github.com/timesler/facenet-pytorch/blob/master/examples/face_tracking.ipynb
", "rawMarkdown": "Next thing to implement: adding a face detection layer => https://github.com/timesler/facenet-pytorch/blob/master/examples/face_tracking.ipynb", "votes": null }, { "id": "758534", "postDate": "02/27/2020 22:06:45", "content": "It works pretty well! Here is a screenshot from one example (you need to imagine this moving through time):
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F172860%2F097eeb0d6e11c84d9d6a275eafc26914%2FScreenshot%20from%202020-02-27%2023-08-00.png?generation=1582841306466224&alt=media\")
Just FYI, the test videos are taken from subdirs 0, 1 and 2 in the training set.
", "votes": null, "replies": [] }, { "id": 747648, "author_name": "pedromoya", "author_url": "", "post_date": "02/16/2020 17:35:39", "content": "thank you! @Yassine Alouini
", "votes": null, "replies": [ { "id": 753547, "author_name": "yassinealouini", "author_url": "", "post_date": "02/22/2020 10:50:07", "content": "Glad it helps!
", "votes": null, "replies": [] } ] }, { "id": 748346, "author_name": "anshulrai", "author_url": "", "post_date": "02/17/2020 12:19:50", "content": "Thank you for sharing @yassinealouini ! I had a couple of questions; what do you mean by audio's number of frames and how is that different from video's number of frames?
Hey @anshulrai. To be honest, I am not sure what these quantities represent yet. \nFor now, I have quickly explored the great metadata dataset from @zaharch\nhttps://www.kaggle.com/zaharch/train-set-metadata-for-dfdc.\nI will update my exploration with more information once I have a better understanding.\nStay tuned. ;)
", "votes": null, "replies": [] } ] }, { "id": 753546, "author_name": "yassinealouini", "author_url": "", "post_date": "02/22/2020 10:49:55", "content": "[EDIT] I was convinced that the source dataset was linked in this post but it wasn't, problem fixed now. :)
", "votes": null, "replies": [] }, { "id": 756574, "author_name": "yassinealouini", "author_url": "", "post_date": "02/25/2020 21:12:54", "content": "[UPDATE] A WIP Streamlit app. More to come in the next few days. ;)
\n\n
Yet more progress, I really like the concept of sidebar in the streamlit API. So far, the API is clean and concise.
\n\n
interesting, I did something similar with streamlit.\nmy app is based on prediction error, some of true/fake are impossible to tell visually
\n\n
Awesome work! Do you mind sharing some details about the model you have there: on how many samples have you trained it, what model architecture, for how long, and so on?\nThanks in advance!
", "votes": null, "replies": [] } ] }, { "id": 758518, "author_name": "yassinealouini", "author_url": "", "post_date": "02/27/2020 21:16:55", "content": "Next thing to implement: adding a face detection layer => https://github.com/timesler/facenet-pytorch/blob/master/examples/face_tracking.ipynb
", "votes": null, "replies": [ { "id": 758534, "author_name": "yassinealouini", "author_url": "", "post_date": "02/27/2020 22:06:45", "content": "It works pretty well! Here is a screenshot from one example (you need to imagine this moving through time):
\n\n