In case anyone is planning on doing any more research on this dataset, I would like to share a manually labelled actor database I built for the real videos.\nEach actor (386 of them) has a unique ID consisting of a pair of numbers. The first is the folder number from training set, and the second is a relative index within that folder of the actor.\nI only labelled the main actor of each video. There are some actors which appear in multiple folders. As far as I can tell, these actors that appear in multiple folders, only appear in multi-person videos of the main actor. So, as far as I know, the actor IDs I labelled only appear in a single folder (in single person videos). This is one reason why I ignored multi-person videos during training. The other reason why I ignored multi-person videos during training was because I never found a single multi-person video where more than one person appeared to be fake in it.
\n\nThe notebook is in https://www.kaggle.com/cesb45/deepfake-detection-actors
\n\nIt contains the dataset, a sample photo of each actor and at the end of the notebook I also have 6 samples of face swaps which if I found the source actor correctly, show how mixed the source and target actors are in the training dataset, such that it seems impossible to find a good validation set of folders whose actors (source and target) are not in the training set. In my opinion, this was a significant weakness of the dataset.
", "messages": [ { "id": 827079, "postDate": "2020-04-30T04:12:04.183Z", "content": "In case anyone is planning on doing any more research on this dataset, I would like to share a manually labelled actor database I built for the real videos.\nEach actor (386 of them) has a unique ID consisting of a pair of numbers. The first is the folder number from training set, and the second is a relative index within that folder of the actor.\nI only labelled the main actor of each video. There are some actors which appear in multiple folders. As far as I can tell, these actors that appear in multiple folders, only appear in multi-person videos of the main actor. So, as far as I know, the actor IDs I labelled only appear in a single folder (in single person videos). This is one reason why I ignored multi-person videos during training. The other reason why I ignored multi-person videos during training was because I never found a single multi-person video where more than one person appeared to be fake in it.
\n\nThe notebook is in https://www.kaggle.com/cesb45/deepfake-detection-actors
\n\nIt contains the dataset, a sample photo of each actor and at the end of the notebook I also have 6 samples of face swaps which if I found the source actor correctly, show how mixed the source and target actors are in the training dataset, such that it seems impossible to find a good validation set of folders whose actors (source and target) are not in the training set. In my opinion, this was a significant weakness of the dataset.
", "rawMarkdown": "In case anyone is planning on doing any more research on this dataset, I would like to share a manually labelled actor database I built for the real videos.\nEach actor (386 of them) has a unique ID consisting of a pair of numbers. The first is the folder number from training set, and the second is a relative index within that folder of the actor.\nI only labelled the main actor of each video. There are some actors which appear in multiple folders. As far as I can tell, these actors that appear in multiple folders, only appear in multi-person videos of the main actor. So, as far as I know, the actor IDs I labelled only appear in a single folder (in single person videos). This is one reason why I ignored multi-person videos during training. The other reason why I ignored multi-person videos during training was because I never found a single multi-person video where more than one person appeared to be fake in it.\n\nThe notebook is in https://www.kaggle.com/cesb45/deepfake-detection-actors\n\nIt contains the dataset, a sample photo of each actor and at the end of the notebook I also have 6 samples of face swaps which if I found the source actor correctly, show how mixed the source and target actors are in the training dataset, such that it seems impossible to find a good validation set of folders whose actors (source and target) are not in the training set. In my opinion, this was a significant weakness of the dataset.", "votes": 3 } ], "comments": [], "raw_markdown_by_id": { "827079": "In case anyone is planning on doing any more research on this dataset, I would like to share a manually labelled actor database I built for the real videos.\nEach actor (386 of them) has a unique ID consisting of a pair of numbers. The first is the folder number from training set, and the second is a relative index within that folder of the actor.\nI only labelled the main actor of each video. There are some actors which appear in multiple folders. As far as I can tell, these actors that appear in multiple folders, only appear in multi-person videos of the main actor. So, as far as I know, the actor IDs I labelled only appear in a single folder (in single person videos). This is one reason why I ignored multi-person videos during training. The other reason why I ignored multi-person videos during training was because I never found a single multi-person video where more than one person appeared to be fake in it.\n\nThe notebook is in https://www.kaggle.com/cesb45/deepfake-detection-actors\n\nIt contains the dataset, a sample photo of each actor and at the end of the notebook I also have 6 samples of face swaps which if I found the source actor correctly, show how mixed the source and target actors are in the training dataset, such that it seems impossible to find a good validation set of folders whose actors (source and target) are not in the training set. In my opinion, this was a significant weakness of the dataset." } }