![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F515279%2F30e66c2f0bc98eb8b0e63d71e322d59c%2FFigure_1.png?generation=1659204666033584&alt=media\")
I read some top-team solutions in the discussion but can't find anyone who used IMU data and improved their score. Is IMU data useless for localization (I don't think so), or if it is the dataset we used in this competition noisy, or if it is just we didn't find a good way to use it? What do you think?
", "messages": [ { "id": "1877479", "postDate": "07/30/2022 17:12:27", "content": "I read some top-team solutions in the discussion but can't find anyone who used IMU data and improved their score. Is IMU data useless for localization (I don't think so), or if it is the dataset we used in this competition noisy, or if it is just we didn't find a good way to use it? What do you think?
", "rawMarkdown": "I read some top-team solutions in the discussion but can't find anyone who used IMU data and improved their score. Is IMU data useless for localization (I don't think so), or if it is the dataset we used in this competition noisy, or if it is just we didn't find a good way to use it? What do you think?", "votes": null }, { "id": "1877537", "postDate": "07/30/2022 18:36:18", "content": "IMU is very important in my opinion
\nSome tracks GYRO contain strong floating bias, but in most case you can get very precise moving dirrection just from baseline and GYRO.Y
\nHere is an example for first track
blue is direction from baseline prediction, green is from gt and yellow direct signal from gyro - gyro match gt a lot, so you can filter baseline and get good initial approximation without outliers
\nThe accelerometer data is more tricky,
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F515279%2F7b671ba2fdf564fb3e0026b2f5a77ad5%2Faccel_048.png?generation=1659205335016155&alt=media\")
Here from down to up - X data from accelerometer, X data from predicted track and X from GT
\nNext Y then Z and then difference between predicted track and actual accel data
\nYou can see, that all - GT, predicted and accelerometer match, but difference is not ideal and I do not know the source of error
\nMore over around 1750 epoch turbulence is too large and in such places track guided by accelerometer have large prediction error
But in most track you can use only IMU data + psevdodistance to get track below 2 meters error
", "rawMarkdown": "IMU is very important in my opinion\nSome tracks GYRO contain strong floating bias, but in most case you can get very precise moving dirrection just from baseline and GYRO.Y\nHere is an example for first track\n\n\n\nblue is direction from baseline prediction, green is from gt and yellow direct signal from gyro - gyro match gt a lot, so you can filter baseline and get good initial approximation without outliers\n\nThe accelerometer data is more tricky, \n\n\n\nHere from down to up - X data from accelerometer, X data from predicted track and X from GT\nNext Y then Z and then difference between predicted track and actual accel data\nYou can see, that all - GT, predicted and accelerometer match, but difference is not ideal and I do not know the source of error\nMore over around 1750 epoch turbulence is too large and in such places track guided by accelerometer have large prediction error\n\nBut in most track you can use only IMU data + psevdodistance to get track below 2 meters error", "votes": null } ], "comments": [ { "id": 1877537, "author_name": "ielenik", "author_url": "", "post_date": "07/30/2022 18:36:18", "content": "IMU is very important in my opinion
\nSome tracks GYRO contain strong floating bias, but in most case you can get very precise moving dirrection just from baseline and GYRO.Y
\nHere is an example for first track
blue is direction from baseline prediction, green is from gt and yellow direct signal from gyro - gyro match gt a lot, so you can filter baseline and get good initial approximation without outliers
\nThe accelerometer data is more tricky,
\n
Here from down to up - X data from accelerometer, X data from predicted track and X from GT
\nNext Y then Z and then difference between predicted track and actual accel data
\nYou can see, that all - GT, predicted and accelerometer match, but difference is not ideal and I do not know the source of error
\nMore over around 1750 epoch turbulence is too large and in such places track guided by accelerometer have large prediction error
But in most track you can use only IMU data + psevdodistance to get track below 2 meters error
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1877479": "I read some top-team solutions in the discussion but can't find anyone who used IMU data and improved their score. Is IMU data useless for localization (I don't think so), or if it is the dataset we used in this competition noisy, or if it is just we didn't find a good way to use it? What do you think?", "1877537": "IMU is very important in my opinion\nSome tracks GYRO contain strong floating bias, but in most case you can get very precise moving dirrection just from baseline and GYRO.Y\nHere is an example for first track\n\n\n\nblue is direction from baseline prediction, green is from gt and yellow direct signal from gyro - gyro match gt a lot, so you can filter baseline and get good initial approximation without outliers\n\nThe accelerometer data is more tricky, \n\n\n\nHere from down to up - X data from accelerometer, X data from predicted track and X from GT\nNext Y then Z and then difference between predicted track and actual accel data\nYou can see, that all - GT, predicted and accelerometer match, but difference is not ideal and I do not know the source of error\nMore over around 1750 epoch turbulence is too large and in such places track guided by accelerometer have large prediction error\n\nBut in most track you can use only IMU data + psevdodistance to get track below 2 meters error" }, "source": "meta" }