Hi guys! \nThe frames from videos present in training dataset are typically of dimensions 1920*1080 pixels which makes the size of 300 such frames together equal to 600 MBs. To what extent can we rescale these frames to reduce memory requirement without losing much on image quality for an optimum model performance?
", "messages": [ { "id": "721622", "postDate": "01/17/2020 14:44:39", "content": "Hi guys! \nThe frames from videos present in training dataset are typically of dimensions 1920*1080 pixels which makes the size of 300 such frames together equal to 600 MBs. To what extent can we rescale these frames to reduce memory requirement without losing much on image quality for an optimum model performance?
", "rawMarkdown": "Hi guys! \nThe frames from videos present in training dataset are typically of dimensions 1920*1080 pixels which makes the size of 300 such frames together equal to 600 MBs. To what extent can we rescale these frames to reduce memory requirement without losing much on image quality for an optimum model performance?", "votes": null }, { "id": "721705", "postDate": "01/17/2020 16:21:49", "content": "I simply save faces, and resize them according to the model(usually 224×224 and 299×299) when training.
", "rawMarkdown": "I simply save faces, and resize them according to the model(usually 224×224 and 299×299) when training.", "votes": null }, { "id": "721861", "postDate": "01/17/2020 19:48:52", "content": "If you want to save space, the best solution is to keep the videos as mp4 files and create a data loader that can read the mp4 file when you need it for training.
\n\nIf you're taking face crops, you can save the bounding box and other metadata in a CSV file. In your data loader you then read from the video and crop the face on-the-fly using the bounding box coordinates from the CSV file.
", "rawMarkdown": "If you want to save space, the best solution is to keep the videos as mp4 files and create a data loader that can read the mp4 file when you need it for training.\n\nIf you're taking face crops, you can save the bounding box and other metadata in a CSV file. In your data loader you then read from the video and crop the face on-the-fly using the bounding box coordinates from the CSV file.", "votes": null }, { "id": "724605", "postDate": "01/21/2020 09:37:45", "content": "Which library do you use ? Is it OpenCV?
", "rawMarkdown": "Which library do you use ? Is it OpenCV?", "votes": null }, { "id": "724737", "postDate": "01/21/2020 12:47:10", "content": "I use OpenCV for reading the videos and BlazeFace for face detection: https://www.kaggle.com/humananalog/starter-blazeface-pytorch
", "rawMarkdown": "I use OpenCV for reading the videos and BlazeFace for face detection: https://www.kaggle.com/humananalog/starter-blazeface-pytorch", "votes": null }, { "id": "725505", "postDate": "01/22/2020 07:28:27", "content": "oh okay. Thanks, I will try it next :)
", "rawMarkdown": "oh okay. Thanks, I will try it next :)", "votes": null }, { "id": "727069", "postDate": "01/23/2020 12:40:06", "content": "@feifeizaici I am new entrant to the competition just got freed.\nCould you please tell the approach here . I so far find multiple ones. \nIn this post are you trying to suggest you generated the Face Images for from Fake and Real videos and then used those static images for classification ?
", "rawMarkdown": "feifeizaici I am new entrant to the competition just got freed.\nCould you please tell the approach here . I so far find multiple ones. \nIn this post are you trying to suggest you generated the Face Images for from Fake and Real videos and then used those static images for classification ?", "votes": null }, { "id": "727223", "postDate": "01/23/2020 14:41:47", "content": "@jaideepvalani Yes, I simply use static images by now. Still trying other methods!
", "rawMarkdown": "jaideepvalani Yes, I simply use static images by now. Still trying other methods!", "votes": null } ], "comments": [ { "id": 721705, "author_name": "feifeizaici", "author_url": "", "post_date": "01/17/2020 16:21:49", "content": "I simply save faces, and resize them according to the model(usually 224×224 and 299×299) when training.
", "votes": null, "replies": [ { "id": 727069, "author_name": "jaideepvalani", "author_url": "", "post_date": "01/23/2020 12:40:06", "content": "@feifeizaici I am new entrant to the competition just got freed.\nCould you please tell the approach here . I so far find multiple ones. \nIn this post are you trying to suggest you generated the Face Images for from Fake and Real videos and then used those static images for classification ?
", "votes": null, "replies": [] }, { "id": 727223, "author_name": "feifeizaici", "author_url": "", "post_date": "01/23/2020 14:41:47", "content": "@jaideepvalani Yes, I simply use static images by now. Still trying other methods!
", "votes": null, "replies": [] } ] }, { "id": 721861, "author_name": "humananalog", "author_url": "", "post_date": "01/17/2020 19:48:52", "content": "If you want to save space, the best solution is to keep the videos as mp4 files and create a data loader that can read the mp4 file when you need it for training.
\n\nIf you're taking face crops, you can save the bounding box and other metadata in a CSV file. In your data loader you then read from the video and crop the face on-the-fly using the bounding box coordinates from the CSV file.
", "votes": null, "replies": [ { "id": 724605, "author_name": "mrunalnasery", "author_url": "", "post_date": "01/21/2020 09:37:45", "content": "Which library do you use ? Is it OpenCV?
", "votes": null, "replies": [] }, { "id": 724737, "author_name": "humananalog", "author_url": "", "post_date": "01/21/2020 12:47:10", "content": "I use OpenCV for reading the videos and BlazeFace for face detection: https://www.kaggle.com/humananalog/starter-blazeface-pytorch
", "votes": null, "replies": [] }, { "id": 725505, "author_name": "mrunalnasery", "author_url": "", "post_date": "01/22/2020 07:28:27", "content": "oh okay. Thanks, I will try it next :)
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "721622": "Hi guys! \nThe frames from videos present in training dataset are typically of dimensions 1920*1080 pixels which makes the size of 300 such frames together equal to 600 MBs. To what extent can we rescale these frames to reduce memory requirement without losing much on image quality for an optimum model performance?", "721705": "I simply save faces, and resize them according to the model(usually 224×224 and 299×299) when training.", "721861": "If you want to save space, the best solution is to keep the videos as mp4 files and create a data loader that can read the mp4 file when you need it for training.\n\nIf you're taking face crops, you can save the bounding box and other metadata in a CSV file. In your data loader you then read from the video and crop the face on-the-fly using the bounding box coordinates from the CSV file.", "724605": "Which library do you use ? Is it OpenCV?", "724737": "I use OpenCV for reading the videos and BlazeFace for face detection: https://www.kaggle.com/humananalog/starter-blazeface-pytorch", "725505": "oh okay. Thanks, I will try it next :)", "727069": "feifeizaici I am new entrant to the competition just got freed.\nCould you please tell the approach here . I so far find multiple ones. \nIn this post are you trying to suggest you generated the Face Images for from Fake and Real videos and then used those static images for classification ?", "727223": "jaideepvalani Yes, I simply use static images by now. Still trying other methods!" }, "source": "meta" }