This dataset is a bit unwieldy if you are making a video-based model. I attempted to extract the faces and compress the cropped videos down to a manageable size. See https://www.kaggle.com/anlthms/dfdc-video-faces for examples.
\n\nI know it is kinda late in the timeline, but does this sound like a reasonable approach to start with? The underlying assumption is that faces are the only relevant part in the videos. One could compress the data down to less than 3% of its size by excluding everything else.
\n\nEDIT: published a notebook to show the face video clips - https://www.kaggle.com/anlthms/example-video-clips
", "messages": [ { "id": "761010", "postDate": "03/02/2020 03:00:06", "content": "This dataset is a bit unwieldy if you are making a video-based model. I attempted to extract the faces and compress the cropped videos down to a manageable size. See https://www.kaggle.com/anlthms/dfdc-video-faces for examples.
\n\nI know it is kinda late in the timeline, but does this sound like a reasonable approach to start with? The underlying assumption is that faces are the only relevant part in the videos. One could compress the data down to less than 3% of its size by excluding everything else.
\n\nEDIT: published a notebook to show the face video clips - https://www.kaggle.com/anlthms/example-video-clips
", "rawMarkdown": "This dataset is a bit unwieldy if you are making a video-based model. I attempted to extract the faces and compress the cropped videos down to a manageable size. See [https://www.kaggle.com/anlthms/dfdc-video-faces](https://www.kaggle.com/anlthms/dfdc-video-faces) for examples.\n\nI know it is kinda late in the timeline, but does this sound like a reasonable approach to start with? The underlying assumption is that faces are the only relevant part in the videos. One could compress the data down to less than 3% of its size by excluding everything else.\n\nEDIT: published a notebook to show the face video clips - [https://www.kaggle.com/anlthms/example-video-clips](https://www.kaggle.com/anlthms/example-video-clips)", "votes": null }, { "id": "761036", "postDate": "03/02/2020 04:04:26", "content": "Yes. And this is the approach currently being used by everyone. Welcome to the competition.
", "rawMarkdown": "Yes. And this is the approach currently being used by everyone. Welcome to the competition.", "votes": null }, { "id": "761074", "postDate": "03/02/2020 05:10:10", "content": "Thanks, @khahuras. If there are example notebooks using video (and not images), I missed them. I am trying to keep the data in video form maintaining the original resolution. The tricky part was motion stabilizing the videos. Has someone already done this?
\n\nPublished a kernel to demo my dataset - https://www.kaggle.com/anlthms/example-video-clips
", "rawMarkdown": "Thanks, @khahuras. If there are example notebooks using video (and not images), I missed them. I am trying to keep the data in video form maintaining the original resolution. The tricky part was motion stabilizing the videos. Has someone already done this?\n\nPublished a kernel to demo my dataset - [https://www.kaggle.com/anlthms/example-video-clips](https://www.kaggle.com/anlthms/example-video-clips)", "votes": null }, { "id": "761102", "postDate": "03/02/2020 06:18:54", "content": "No, at least no public thing about that to my knowledge.
", "rawMarkdown": "No, at least no public thing about that to my knowledge.", "votes": null }, { "id": "761124", "postDate": "03/02/2020 06:59:10", "content": "@anlthms Do you mean this one. I've almost done it.\nhttps://www.kaggle.com/wuliaokaola/dfdc-face-videos-part01
", "rawMarkdown": "anlthms Do you mean this one. I've almost done it.\nhttps://www.kaggle.com/wuliaokaola/dfdc-face-videos-part01", "votes": null }, { "id": "761626", "postDate": "03/02/2020 19:18:26", "content": "Yes, your dataset looks similar. Did it help to keep the input data as video?
\n\nMy intuition is that image-based models will miss out on features in the time dimension. Things like flicker artifacts would be easier to detect with temporal data.
", "rawMarkdown": "Yes, your dataset looks similar. Did it help to keep the input data as video?\n\nMy intuition is that image-based models will miss out on features in the time dimension. Things like flicker artifacts would be easier to detect with temporal data.", "votes": null }, { "id": "761803", "postDate": "03/03/2020 00:23:01", "content": "I plan to implement the time series model. But now I am working on image-based model. The current LB score (0.39692) is based on a single image-based model without ensemble.
\n\nIf anyone want to team up, please PM me. I have both AWS credit and GCP TPU quota.
", "rawMarkdown": "I plan to implement the time series model. But now I am working on image-based model. The current LB score (0.39692) is based on a single image-based model without ensemble.\n\nIf anyone want to team up, please PM me. I have both AWS credit and GCP TPU quota.", "votes": null }, { "id": "764747", "postDate": "03/05/2020 20:39:17", "content": "I put together some minimalistic code to train and validate using these cropped videos.
\n\nhttps://www.kaggle.com/anlthms/video-classification-with-3d-resnet
\n\nUsing 3 folders to train and 2 to validate:
\n\n\n\n\nvalidation accuracy 56%\n epoch 6 training loss 0.63 validation loss 0.68
\n
Was hoping for better results. Oh well...
", "rawMarkdown": "I put together some minimalistic code to train and validate using these cropped videos. \n\n[https://www.kaggle.com/anlthms/video-classification-with-3d-resnet](https://www.kaggle.com/anlthms/video-classification-with-3d-resnet)\n\nUsing 3 folders to train and 2 to validate:\n\n> validation accuracy 56%\n> epoch 6 training loss 0.63 validation loss 0.68\n\nWas hoping for better results. Oh well...", "votes": null } ], "comments": [ { "id": 761036, "author_name": "khahuras", "author_url": "", "post_date": "03/02/2020 04:04:26", "content": "Yes. And this is the approach currently being used by everyone. Welcome to the competition.
", "votes": null, "replies": [ { "id": 761074, "author_name": "anlthms", "author_url": "", "post_date": "03/02/2020 05:10:10", "content": "Thanks, @khahuras. If there are example notebooks using video (and not images), I missed them. I am trying to keep the data in video form maintaining the original resolution. The tricky part was motion stabilizing the videos. Has someone already done this?
\n\nPublished a kernel to demo my dataset - https://www.kaggle.com/anlthms/example-video-clips
", "votes": null, "replies": [] }, { "id": 761102, "author_name": "khahuras", "author_url": "", "post_date": "03/02/2020 06:18:54", "content": "No, at least no public thing about that to my knowledge.
", "votes": null, "replies": [] }, { "id": 761124, "author_name": "wuliaokaola", "author_url": "", "post_date": "03/02/2020 06:59:10", "content": "@anlthms Do you mean this one. I've almost done it.\nhttps://www.kaggle.com/wuliaokaola/dfdc-face-videos-part01
", "votes": null, "replies": [] }, { "id": 761626, "author_name": "anlthms", "author_url": "", "post_date": "03/02/2020 19:18:26", "content": "Yes, your dataset looks similar. Did it help to keep the input data as video?
\n\nMy intuition is that image-based models will miss out on features in the time dimension. Things like flicker artifacts would be easier to detect with temporal data.
", "votes": null, "replies": [] }, { "id": 761803, "author_name": "wuliaokaola", "author_url": "", "post_date": "03/03/2020 00:23:01", "content": "I plan to implement the time series model. But now I am working on image-based model. The current LB score (0.39692) is based on a single image-based model without ensemble.
\n\nIf anyone want to team up, please PM me. I have both AWS credit and GCP TPU quota.
", "votes": null, "replies": [] } ] }, { "id": 764747, "author_name": "anlthms", "author_url": "", "post_date": "03/05/2020 20:39:17", "content": "I put together some minimalistic code to train and validate using these cropped videos.
\n\nhttps://www.kaggle.com/anlthms/video-classification-with-3d-resnet
\n\nUsing 3 folders to train and 2 to validate:
\n\n\n\n\nvalidation accuracy 56%\n epoch 6 training loss 0.63 validation loss 0.68
\n
Was hoping for better results. Oh well...
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "761010": "This dataset is a bit unwieldy if you are making a video-based model. I attempted to extract the faces and compress the cropped videos down to a manageable size. See [https://www.kaggle.com/anlthms/dfdc-video-faces](https://www.kaggle.com/anlthms/dfdc-video-faces) for examples.\n\nI know it is kinda late in the timeline, but does this sound like a reasonable approach to start with? The underlying assumption is that faces are the only relevant part in the videos. One could compress the data down to less than 3% of its size by excluding everything else.\n\nEDIT: published a notebook to show the face video clips - [https://www.kaggle.com/anlthms/example-video-clips](https://www.kaggle.com/anlthms/example-video-clips)", "761036": "Yes. And this is the approach currently being used by everyone. Welcome to the competition.", "761074": "Thanks, @khahuras. If there are example notebooks using video (and not images), I missed them. I am trying to keep the data in video form maintaining the original resolution. The tricky part was motion stabilizing the videos. Has someone already done this?\n\nPublished a kernel to demo my dataset - [https://www.kaggle.com/anlthms/example-video-clips](https://www.kaggle.com/anlthms/example-video-clips)", "761102": "No, at least no public thing about that to my knowledge.", "761124": "anlthms Do you mean this one. I've almost done it.\nhttps://www.kaggle.com/wuliaokaola/dfdc-face-videos-part01", "761626": "Yes, your dataset looks similar. Did it help to keep the input data as video?\n\nMy intuition is that image-based models will miss out on features in the time dimension. Things like flicker artifacts would be easier to detect with temporal data.", "761803": "I plan to implement the time series model. But now I am working on image-based model. The current LB score (0.39692) is based on a single image-based model without ensemble.\n\nIf anyone want to team up, please PM me. I have both AWS credit and GCP TPU quota.", "764747": "I put together some minimalistic code to train and validate using these cropped videos. \n\n[https://www.kaggle.com/anlthms/video-classification-with-3d-resnet](https://www.kaggle.com/anlthms/video-classification-with-3d-resnet)\n\nUsing 3 folders to train and 2 to validate:\n\n> validation accuracy 56%\n> epoch 6 training loss 0.63 validation loss 0.68\n\nWas hoping for better results. Oh well..." }, "source": "meta" }