\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4310004%2Ff33528fc08d922a2bf72cb071405815a%2F16179447-kg.png?generation=1753088705928282&alt=media\")
Just wondering how much CV and LB improvement you can achieve with feature engineering (imu only).
", "messages": [ { "id": 3251609, "postDate": "2025-07-21T07:40:58.180Z", "content": "After exploring more deeply, I still can't believe that providing only acceleration and quaternion data can achieve so much for recognizing gestures.
\nI do a kind of processing on acc like this:
\n
and it shows good improvement for judging the gestures
", "rawMarkdown": "After exploring more deeply, I still can't believe that providing only acceleration and quaternion data can achieve so much for recognizing gestures.\n\nI do a kind of processing on acc like this:\n\n\nand it shows good improvement for judging the gestures", "votes": 3, "replies": [ { "id": 3251799, "postDate": "2025-07-21T15:25:35.130Z", "content": "May i ask what the blue points represent?
", "rawMarkdown": "May i ask what the blue points represent? " }, { "id": 3251920, "postDate": "2025-07-21T18:48:11.113Z", "content": "Great! I also do a processing like this and it shows good improvement ![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4629460%2Ffe0427f6d7288eb7ba91cad1c7017485%2FSteamgraph-1024x962.png?generation=1753123646509605&alt=media\")
Would you please explain what this is? It almost looks like augmentations of transition phase of a sequence but there’s a bit going on without much context
", "rawMarkdown": "Would you please explain what this is? It almost looks like augmentations of transition phase of a sequence but there’s a bit going on without much context", "replies": [ { "id": 3252053, "postDate": "2025-07-22T00:14:53.773Z", "content": "@jackvd this is an algorithm applied on the acc, the blue point you see on the plot is original acc. It's similar to my MCMC notebook you saw before. But more advanced.
", "rawMarkdown": "@jackvd this is an algorithm applied on the acc, the blue point you see on the plot is original acc. It's similar to my MCMC notebook you saw before. But more advanced." } ] } ] }, { "id": 3251334, "postDate": "2025-07-20T13:55:45.393Z", "content": "Just wondering how much CV and LB improvement you can achieve with feature engineering (imu only).
", "rawMarkdown": "Just wondering how much CV and LB improvement you can achieve with feature engineering (imu only).", "votes": 4 }, { "id": 3252297, "postDate": "2025-07-22T12:06:53.017Z", "content": "Depends what you mean by \"improve.\"
\nIf you're squeezing signal from IMU, then sure—custom features (peak intervals, entropy metrics, etc.) will bump your CV a few notches. But I noticed in these kinds of datasets, there's this strange semantic ceiling: you can handcraft 50 clever features, but once the noise entangles with latent patterns, they stop helping.
\nIn my case, I started asking: what if the failure isn’t in engineering, but in how the model understands intention? Like, the real bottleneck might be the structure of reasoning, not the features.
\nI wrote a short breakdown about this frustration—why some datasets seem to resist feature extraction, no matter how much you tweak. It’s open source. Strangely, even the author of tesseract.js (OCR legend) starred it, which surprised me:
\nhttps://github.com/onestardao/WFGY
Not selling anything—just sharing in case you’ve ever hit that wall where the model’s \"improvement\" feels… semantic, not numeric. Curious if anyone else felt the same.
", "rawMarkdown": "Depends what you mean by \"improve.\"\n\nIf you're squeezing signal from IMU, then sure—custom features (peak intervals, entropy metrics, etc.) will bump your CV a few notches. But I noticed in these kinds of datasets, there's this strange semantic ceiling: you can handcraft 50 clever features, but once the noise entangles with latent patterns, they stop helping.\n\nIn my case, I started asking: what if the failure isn’t in engineering, but in how the model understands intention? Like, the real bottleneck might be the structure of reasoning, not the features.\n\nI wrote a short breakdown about this frustration—why some datasets seem to resist feature extraction, no matter how much you tweak. It’s open source. Strangely, even the author of tesseract.js (OCR legend) starred it, which surprised me:\nhttps://github.com/onestardao/WFGY\n\nNot selling anything—just sharing in case you’ve ever hit that wall where the model’s \"improvement\" feels... semantic, not numeric. Curious if anyone else felt the same.", "votes": -8 }, { "id": 3252068, "postDate": "2025-07-22T01:14:21.253Z", "content": "1d-cnn: cv0.7779 lb:0.79(high) -> cv0.7930 lb:0.81(mid)
", "rawMarkdown": "1d-cnn: cv0.7779 lb:0.79(high) -> cv0.7930 lb:0.81(mid)" }, { "id": 3251569, "postDate": "2025-07-21T05:59:42.803Z", "content": "imu only
\ncv:0.75~0.77
\nlb:0.76~0.79
After exploring more deeply, I still can't believe that providing only acceleration and quaternion data can achieve so much for recognizing gestures.
\nI do a kind of processing on acc like this:
\n
and it shows good improvement for judging the gestures
", "votes": 3, "replies": [ { "id": 3251799, "author_name": "Gerwyn", "author_url": "", "post_date": "2025-07-21T15:25:35.130000", "content": "May i ask what the blue points represent?
", "votes": 0, "replies": [] }, { "id": 3251920, "author_name": "Cihat Emre Çeliker", "author_url": "", "post_date": "2025-07-21T18:48:11.113000", "content": "Great! I also do a processing like this and it shows good improvement
Would you please explain what this is? It almost looks like augmentations of transition phase of a sequence but there’s a bit going on without much context
", "votes": 0, "replies": [ { "id": 3252053, "author_name": "Tom", "author_url": "", "post_date": "2025-07-22T00:14:53.773000", "content": "@jackvd this is an algorithm applied on the acc, the blue point you see on the plot is original acc. It's similar to my MCMC notebook you saw before. But more advanced.
", "votes": 0, "replies": [] } ] } ] }, { "id": 3252297, "author_name": "PSBigBig", "author_url": "", "post_date": "2025-07-22T12:06:53.017000", "content": "Depends what you mean by \"improve.\"
\nIf you're squeezing signal from IMU, then sure—custom features (peak intervals, entropy metrics, etc.) will bump your CV a few notches. But I noticed in these kinds of datasets, there's this strange semantic ceiling: you can handcraft 50 clever features, but once the noise entangles with latent patterns, they stop helping.
\nIn my case, I started asking: what if the failure isn’t in engineering, but in how the model understands intention? Like, the real bottleneck might be the structure of reasoning, not the features.
\nI wrote a short breakdown about this frustration—why some datasets seem to resist feature extraction, no matter how much you tweak. It’s open source. Strangely, even the author of tesseract.js (OCR legend) starred it, which surprised me:
\nhttps://github.com/onestardao/WFGY
Not selling anything—just sharing in case you’ve ever hit that wall where the model’s \"improvement\" feels… semantic, not numeric. Curious if anyone else felt the same.
", "votes": -8, "replies": [] }, { "id": 3252068, "author_name": "Kurise", "author_url": "", "post_date": "2025-07-22T01:14:21.253000", "content": "1d-cnn: cv0.7779 lb:0.79(high) -> cv0.7930 lb:0.81(mid)
", "votes": 0, "replies": [] }, { "id": 3251569, "author_name": "XuKong Ji", "author_url": "", "post_date": "2025-07-21T05:59:42.803000", "content": "imu only
\ncv:0.75~0.77
\nlb:0.76~0.79