Following the release of the 1 million fake faces and the 70k real flickr faces datasets, I was wondering if the kaggle community would be interested the simpler (\"toy\") task of training a model that can correctly classify an image as being real (i.e. comes from the real flickr dataset), or fake (generated by StyleGAN).
\n\nSpecifically, I already wrote some code to convert all the high-quality images into jpeg-compressed 256x256 images, so that they can be easily downloaded, and makes the classification task slightly more realistic (most real-life images will have some sort of compression loss). I sampled 70k fake images from Bojan's dataset, and used all 70k flikr images, for a total size of 140k images; this is subsequently split into 100k in the training set, 20k in validation set, 20k in test set.
\n\nI was wondering: would you prefer if I created a dataset containing all this data, or a \"in-class\" competition (open to public) where the test set would be released at the end?
\n\nThe latter would obviously not have prizes, but it could be a nice opportunity to compare different ideas by leveraging the scoring system. I could also spice it up by adding extra challenges to the test set that would reflect real-life scenarios (gaussian noise, augmentations, more compression artifact, segment-and-blur background, etc.).
\n\nI have never hosted a competition before, but if the community likes this idea it's something I'd love to do!
", "messages": [ { "id": "740674", "postDate": "02/09/2020 16:37:16", "content": "Following the release of the 1 million fake faces and the 70k real flickr faces datasets, I was wondering if the kaggle community would be interested the simpler (\"toy\") task of training a model that can correctly classify an image as being real (i.e. comes from the real flickr dataset), or fake (generated by StyleGAN).
\n\nSpecifically, I already wrote some code to convert all the high-quality images into jpeg-compressed 256x256 images, so that they can be easily downloaded, and makes the classification task slightly more realistic (most real-life images will have some sort of compression loss). I sampled 70k fake images from Bojan's dataset, and used all 70k flikr images, for a total size of 140k images; this is subsequently split into 100k in the training set, 20k in validation set, 20k in test set.
\n\nI was wondering: would you prefer if I created a dataset containing all this data, or a \"in-class\" competition (open to public) where the test set would be released at the end?
\n\nThe latter would obviously not have prizes, but it could be a nice opportunity to compare different ideas by leveraging the scoring system. I could also spice it up by adding extra challenges to the test set that would reflect real-life scenarios (gaussian noise, augmentations, more compression artifact, segment-and-blur background, etc.).
\n\nI have never hosted a competition before, but if the community likes this idea it's something I'd love to do!
", "rawMarkdown": "Following the release of the [1 million fake faces](https://www.kaggle.com/c/deepfake-detection-challenge/discussion/121173) and the [70k real flickr faces](https://www.kaggle.com/c/deepfake-detection-challenge/discussion/122786) datasets, I was wondering if the kaggle community would be interested **the simpler (\"toy\") task of training a model that can correctly classify an image as being real** (i.e. comes from the real flickr dataset), or fake (generated by StyleGAN). \n\nSpecifically, I already wrote some code to convert all the high-quality images into jpeg-compressed 256x256 images, so that they can be easily downloaded, and makes the classification task slightly more realistic (most real-life images will have some sort of compression loss). I sampled 70k fake images from Bojan's dataset, and used all 70k flikr images, for a total size of 140k images; this is subsequently split into 100k in the training set, 20k in validation set, 20k in test set.\n\nI was wondering: **would you prefer if I created a dataset containing all this data, or a \"in-class\" competition (open to public) where the test set would be released at the end?** \n\nThe latter would obviously not have prizes, but it could be a nice opportunity to compare different ideas by leveraging the scoring system. I could also spice it up by adding extra challenges to the **test set** that would reflect real-life scenarios (gaussian noise, augmentations, more compression artifact, segment-and-blur background, etc.).\n\nI have never hosted a competition before, but if the community likes this idea it's something I'd love to do!", "votes": null }, { "id": "740675", "postDate": "02/09/2020 16:38:50", "content": "Also check-in with @kaggleteam : Do in-class competition need to be hosted only for a closed community or special event, or anyone can host them? Just want to make sure I'm not breaking any rule if I do :)
", "rawMarkdown": "Also check-in with @kaggleteam : Do in-class competition need to be hosted only for a closed community or special event, or anyone can host them? Just want to make sure I'm not breaking any rule if I do :)", "votes": null }, { "id": "743343", "postDate": "02/12/2020 02:03:05", "content": "It'd be great to have a in-class competition. But not sure whether it cost you money.
", "rawMarkdown": "It'd be great to have a in-class competition. But not sure whether it cost you money.", "votes": null }, { "id": "743357", "postDate": "02/12/2020 02:23:16", "content": "No, I think it's completely free. I'm just not sure if it's allowed to host it in a non-class setting.
", "rawMarkdown": "No, I think it's completely free. I'm just not sure if it's allowed to host it in a non-class setting.", "votes": null }, { "id": "743370", "postDate": "02/12/2020 02:31:46", "content": "I'm sure it would be ok.
", "rawMarkdown": "I'm sure it would be ok.", "votes": null } ], "comments": [ { "id": 740675, "author_name": "xhlulu", "author_url": "", "post_date": "02/09/2020 16:38:50", "content": "Also check-in with @kaggleteam : Do in-class competition need to be hosted only for a closed community or special event, or anyone can host them? Just want to make sure I'm not breaking any rule if I do :)
", "votes": null, "replies": [] }, { "id": 743343, "author_name": "unkownhihi", "author_url": "", "post_date": "02/12/2020 02:03:05", "content": "It'd be great to have a in-class competition. But not sure whether it cost you money.
", "votes": null, "replies": [ { "id": 743357, "author_name": "xhlulu", "author_url": "", "post_date": "02/12/2020 02:23:16", "content": "No, I think it's completely free. I'm just not sure if it's allowed to host it in a non-class setting.
", "votes": null, "replies": [] }, { "id": 743370, "author_name": "unkownhihi", "author_url": "", "post_date": "02/12/2020 02:31:46", "content": "I'm sure it would be ok.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "740674": "Following the release of the [1 million fake faces](https://www.kaggle.com/c/deepfake-detection-challenge/discussion/121173) and the [70k real flickr faces](https://www.kaggle.com/c/deepfake-detection-challenge/discussion/122786) datasets, I was wondering if the kaggle community would be interested **the simpler (\"toy\") task of training a model that can correctly classify an image as being real** (i.e. comes from the real flickr dataset), or fake (generated by StyleGAN). \n\nSpecifically, I already wrote some code to convert all the high-quality images into jpeg-compressed 256x256 images, so that they can be easily downloaded, and makes the classification task slightly more realistic (most real-life images will have some sort of compression loss). I sampled 70k fake images from Bojan's dataset, and used all 70k flikr images, for a total size of 140k images; this is subsequently split into 100k in the training set, 20k in validation set, 20k in test set.\n\nI was wondering: **would you prefer if I created a dataset containing all this data, or a \"in-class\" competition (open to public) where the test set would be released at the end?** \n\nThe latter would obviously not have prizes, but it could be a nice opportunity to compare different ideas by leveraging the scoring system. I could also spice it up by adding extra challenges to the **test set** that would reflect real-life scenarios (gaussian noise, augmentations, more compression artifact, segment-and-blur background, etc.).\n\nI have never hosted a competition before, but if the community likes this idea it's something I'd love to do!", "740675": "Also check-in with @kaggleteam : Do in-class competition need to be hosted only for a closed community or special event, or anyone can host them? Just want to make sure I'm not breaking any rule if I do :)", "743343": "It'd be great to have a in-class competition. But not sure whether it cost you money.", "743357": "No, I think it's completely free. I'm just not sure if it's allowed to host it in a non-class setting.", "743370": "I'm sure it would be ok." }, "source": "meta" }