![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F10311266%2F19cda438589b2d2bcbb78dfa6b78ad8e%2Fmost-occurred-label-dist.png?generation=1659367406615919&alt=media\")
We are new to Kaggle competitions, and it's the first time for us to earn silver medal.
\nWe had learned a lot of things related to GNSS analysis for first a month after joined this competition because we didn't know anything about it. It has been quite enjoyable to know how to determine precise positions using GNSS satellites discovered by the receiver.
\nWe would like to thank much the hosts for having this exciting competitions and the participants for sharing great notebooks, especially @taroz1461 and @timeverett.
We will introduce first 3 topics in this discussion. (Skipped the others that look duplicate in the other top-place solutions)
\nWe have investigated and annotated every drive route using Google Earth Pro, following the manner shown in the discussion of the last competition by @taroz1461.
\nThe distribution of these annotations is in the following.
All the satellite-selection criteria is determined for each drive-route annotation in optimization by Optuna.
\nWe thought the satellite-selection criteria should change in different environment situations, so apply optimization for every 3 drive-route annotations.
\nFinally, we got improving public LB score 2.64 from 2.65 a little without kalman smoother (2.132 from 2.140 in private LB score).
But unfortunately, we noticed we can get private score 1.784 with the annotation-independent postprocessing after the competition finished.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F10311266%2F365f821b9687c7063335b88c2378dee9%2Fbest_score_on_late_submission.png?generation=1659365999974204&alt=media\")
Our visualization of drive route tells us that the fluctuations of the estimated positions look significantly large on stopping, particularly in 'Tree' or 'Downtown'.
\nWe simply multiply 3 times of position covariance when the speed of the receiver is less than 1 meter per a second for drive-route annotations except for 'Highway', which led the public score to 2.38 from 2.41 (1.863 from 1.904 in the private score).
We are new to Kaggle competitions, and it's the first time for us to earn silver medal.
\nWe had learned a lot of things related to GNSS analysis for first a month after joined this competition because we didn't know anything about it. It has been quite enjoyable to know how to determine precise positions using GNSS satellites discovered by the receiver.
\nWe would like to thank much the hosts for having this exciting competitions and the participants for sharing great notebooks, especially @taroz1461 and @timeverett.
We will introduce first 3 topics in this discussion. (Skipped the others that look duplicate in the other top-place solutions)
\nWe have investigated and annotated every drive route using Google Earth Pro, following the manner shown in the discussion of the last competition by @taroz1461.
\nThe distribution of these annotations is in the following.
All the satellite-selection criteria is determined for each drive-route annotation in optimization by Optuna.
\nWe thought the satellite-selection criteria should change in different environment situations, so apply optimization for every 3 drive-route annotations.
\nFinally, we got improving public LB score 2.64 from 2.65 a little without kalman smoother (2.132 from 2.140 in private LB score).
But unfortunately, we noticed we can get private score 1.784 with the annotation-independent postprocessing after the competition finished.
\n
Our visualization of drive route tells us that the fluctuations of the estimated positions look significantly large on stopping, particularly in 'Tree' or 'Downtown'.
\nWe simply multiply 3 times of position covariance when the speed of the receiver is less than 1 meter per a second for drive-route annotations except for 'Highway', which led the public score to 2.38 from 2.41 (1.863 from 1.904 in the private score).
Very well explained article, hope this helps all of us trying to learn and apply these methods.
", "rawMarkdown": "Very well explained article, hope this helps all of us trying to learn and apply these methods.", "votes": null } ], "comments": [ { "id": 1882623, "author_name": "ravi20076", "author_url": "", "post_date": "08/03/2022 11:07:08", "content": "Very well explained article, hope this helps all of us trying to learn and apply these methods.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1880284": "We are new to Kaggle competitions, and it's the first time for us to earn silver medal.\nWe had learned a lot of things related to GNSS analysis for first a month after joined this competition because we didn't know anything about it. It has been quite enjoyable to know how to determine precise positions using GNSS satellites discovered by the receiver.\nWe would like to thank much the hosts for having this exciting competitions and the participants for sharing great notebooks, especially @taroz1461 and @timeverett.\n\n# Overview\n\n- Annotating every drive route with one of 'Highway', 'Tree' or 'Downtown'\n- Optimization of Satellite-Selection Criteria by Optuna\n- When Stopping, 3x Multiplication of Estimated-Position Covariance passing into Kalman Smoother\n- Robust WSL to determine Positions and Velocities and Kalman Smoother, provided in [this notebook](https://www.kaggle.com/code/taroz1461/carrier-smoothing-robust-wls-kalman-smoother) by @taroz1461\n- RTKLIB to determine Positions, provided in [this notebook](https://www.kaggle.com/code/timeverett/getting-started-with-rtklib) by @timeverett\n\nWe will introduce first 3 topics in this discussion. (Skipped the others that look duplicate in the other top-place solutions)\n\n## Drive-Route Annotations\n\nWe have investigated and annotated every drive route using Google Earth Pro, following the manner shown in [the discussion](https://www.kaggle.com/competitions/google-smartphone-decimeter-challenge/discussion/245160) of the last competition by @taroz1461.\nThe distribution of these annotations is in the following.\n\n\n\n## Optimization of Satellite-Selection Criteria\n\nAll the satellite-selection criteria is determined for each drive-route annotation in optimization by Optuna.\nWe thought the satellite-selection criteria should change in different environment situations, so apply optimization for every 3 drive-route annotations.\nFinally, we got improving public LB score 2.64 from 2.65 a little without kalman smoother (2.132 from 2.140 in private LB score).\n\nBut unfortunately, we noticed we can get private score 1.784 with the annotation-independent postprocessing after the competition finished.\n\n\n\n## When Stopping, 3x Multiplication of Estimated-Position Covariance\n\nOur visualization of drive route tells us that the fluctuations of the estimated positions look significantly large on stopping, particularly in 'Tree' or 'Downtown'.\nWe simply multiply 3 times of position covariance when the speed of the receiver is less than 1 meter per a second for drive-route annotations except for 'Highway', which led the public score to 2.38 from 2.41 (1.863 from 1.904 in the private score).", "1882623": "Very well explained article, hope this helps all of us trying to learn and apply these methods." }, "source": "meta" }