For those who have done sufficient EDA and are comfortable sharing, how many frames are you analyzing before time_of_alert? I've eye-balled a lot of videos in both train and test and Intuitively, it feels like it shouldnt really be more than about 2 seconds (30fps * 2 = 60 frames) before which, the frames should more or less be independent of the target.
", "messages": [ { "id": "3137252", "postDate": "03/01/2025 10:52:11", "content": "For those who have done sufficient EDA and are comfortable sharing, how many frames are you analyzing before time_of_alert? I've eye-balled a lot of videos in both train and test and Intuitively, it feels like it shouldnt really be more than about 2 seconds (30fps * 2 = 60 frames) before which, the frames should more or less be independent of the target.
", "rawMarkdown": "For those who have done sufficient EDA and are comfortable sharing, how many frames are you analyzing before time_of_alert? I've eye-balled a lot of videos in both train and test and Intuitively, it feels like it shouldnt really be more than about 2 seconds (30fps * 2 = 60 frames) before which, the frames should more or less be independent of the target.", "votes": null }, { "id": "3137374", "postDate": "03/01/2025 14:03:25", "content": "As a driver I had the same thoughs. To take no much more than 2 seconds since as farther from event more irrelevant information. So 64 frames worked well for my first CNN 3D attempts. But just because in DL you never know well till you try I'm taking higher downsampled periods right now and I've got better results. I'm training in 12 GB GPU with kaggle replicates when I find a significant better pipeline.
", "rawMarkdown": "As a driver I had the same thoughs. To take no much more than 2 seconds since as farther from event more irrelevant information. So 64 frames worked well for my first CNN 3D attempts. But just because in DL you never know well till you try I'm taking higher downsampled periods right now and I've got better results. I'm training in 12 GB GPU with kaggle replicates when I find a significant better pipeline.", "votes": null }, { "id": "3137465", "postDate": "03/01/2025 16:13:24", "content": "I would say 2~3 sec before the event is enough for humans to \"feel the accident\". But for models, history context frames further than 2-3 sec from the event may also be useful. For example, model may need the further history context to learn what is normal and what is abnormal. Just some intuitive thoughts. After all, in DL, all you need is experimenting.
", "rawMarkdown": "I would say 2~3 sec before the event is enough for humans to \"feel the accident\". But for models, history context frames further than 2-3 sec from the event may also be useful. For example, model may need the further history context to learn what is normal and what is abnormal. Just some intuitive thoughts. After all, in DL, all you need is experimenting.", "votes": null }, { "id": "3140147", "postDate": "03/04/2025 10:31:52", "content": "But if frames before time_of_alert can be used to predict the outcome, what is even the point of time_of_alert?
", "rawMarkdown": "But if frames before time_of_alert can be used to predict the outcome, what is even the point of time_of_alert?", "votes": null }, { "id": "3140170", "postDate": "03/04/2025 11:07:51", "content": "to let you know if a larger tta can be applied to this video.
", "rawMarkdown": "to let you know if a larger tta can be applied to this video.", "votes": null }, { "id": "3144372", "postDate": "03/08/2025 09:20:15", "content": "Can you expand what you mean by this?
", "rawMarkdown": "Can you expand what you mean by this?", "votes": null } ], "comments": [ { "id": 3137374, "author_name": "sacuscreed", "author_url": "", "post_date": "03/01/2025 14:03:25", "content": "As a driver I had the same thoughs. To take no much more than 2 seconds since as farther from event more irrelevant information. So 64 frames worked well for my first CNN 3D attempts. But just because in DL you never know well till you try I'm taking higher downsampled periods right now and I've got better results. I'm training in 12 GB GPU with kaggle replicates when I find a significant better pipeline.
", "votes": null, "replies": [] }, { "id": 3137465, "author_name": "shiyili", "author_url": "", "post_date": "03/01/2025 16:13:24", "content": "I would say 2~3 sec before the event is enough for humans to \"feel the accident\". But for models, history context frames further than 2-3 sec from the event may also be useful. For example, model may need the further history context to learn what is normal and what is abnormal. Just some intuitive thoughts. After all, in DL, all you need is experimenting.
", "votes": null, "replies": [] }, { "id": 3140147, "author_name": "elhnan", "author_url": "", "post_date": "03/04/2025 10:31:52", "content": "But if frames before time_of_alert can be used to predict the outcome, what is even the point of time_of_alert?
", "votes": null, "replies": [ { "id": 3140170, "author_name": "shiyili", "author_url": "", "post_date": "03/04/2025 11:07:51", "content": "to let you know if a larger tta can be applied to this video.
", "votes": null, "replies": [ { "id": 3144372, "author_name": "elhnan", "author_url": "", "post_date": "03/08/2025 09:20:15", "content": "Can you expand what you mean by this?
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3137252": "For those who have done sufficient EDA and are comfortable sharing, how many frames are you analyzing before time_of_alert? I've eye-balled a lot of videos in both train and test and Intuitively, it feels like it shouldnt really be more than about 2 seconds (30fps * 2 = 60 frames) before which, the frames should more or less be independent of the target.", "3137374": "As a driver I had the same thoughs. To take no much more than 2 seconds since as farther from event more irrelevant information. So 64 frames worked well for my first CNN 3D attempts. But just because in DL you never know well till you try I'm taking higher downsampled periods right now and I've got better results. I'm training in 12 GB GPU with kaggle replicates when I find a significant better pipeline.", "3137465": "I would say 2~3 sec before the event is enough for humans to \"feel the accident\". But for models, history context frames further than 2-3 sec from the event may also be useful. For example, model may need the further history context to learn what is normal and what is abnormal. Just some intuitive thoughts. After all, in DL, all you need is experimenting.", "3140147": "But if frames before time_of_alert can be used to predict the outcome, what is even the point of time_of_alert?", "3140170": "to let you know if a larger tta can be applied to this video.", "3144372": "Can you expand what you mean by this?" }, "source": "meta" }