This competition has only one month left.
\nThe dataset is very complex and I think many teams, including us, are having a hard time. Also, some of the people who have just joined may not know what to start with.
So our team will once again update the questions and answer them like a month ago. Please feel free to reply. Any comments are welcome!
\nQuestion
\nOur answer
\nWe use machine learning as part of our post-processing. It's definitely more accurate than a rule-based approach that doesn't use machine learning. If you're having trouble with post-processing, try to figure out if you can replace the post-processing you're using with machine learning.
Some post-processing is applied separately to SJC and others, and some post-processing is applied only to SJC.
It is very noisy and we have not been able to use it well yet.
We have been working hard since we participated in the competition, but we haven't had a positive impact on LB yet. In other words, our LB were achieved using only post-processing and no domain knowledge at all. This might encourage people who don't have domain knowledge:)
Finally, let's enjoy this competition with all our strength for the remaining one month!
", "messages": [ { "id": 1376476, "postDate": "2021-07-05T06:46:37.467Z", "content": "This competition has only one month left.
\nThe dataset is very complex and I think many teams, including us, are having a hard time. Also, some of the people who have just joined may not know what to start with.
So our team will once again update the questions and answer them like a month ago. Please feel free to reply. Any comments are welcome!
\nQuestion
\nOur answer
\nWe use machine learning as part of our post-processing. It's definitely more accurate than a rule-based approach that doesn't use machine learning. If you're having trouble with post-processing, try to figure out if you can replace the post-processing you're using with machine learning.
Some post-processing is applied separately to SJC and others, and some post-processing is applied only to SJC.
It is very noisy and we have not been able to use it well yet.
We have been working hard since we participated in the competition, but we haven't had a positive impact on LB yet. In other words, our LB were achieved using only post-processing and no domain knowledge at all. This might encourage people who don't have domain knowledge:)
Finally, let's enjoy this competition with all our strength for the remaining one month!
", "rawMarkdown": "This competition has only one month left.\nThe dataset is very complex and I think many teams, including us, are having a hard time. Also, some of the people who have just joined may not know what to start with.\n\nSo our team will once again update the questions and answer them [like a month ago](https://www.kaggle.com/c/google-smartphone-decimeter-challenge/discussion/244752). Please feel free to reply. Any comments are welcome!\n\n**Question**\n1. Are you taking advantage of machine learning?\n2. Do you have SJC-specific post-processing?\n3. Are you using IMU data?\n4. Have you succeeded in estimating the absolute position using RTKLIB etc.?\n\n**Our answer**\n 1. We use machine learning as part of our post-processing. It's definitely more accurate than a rule-based approach that doesn't use machine learning. If you're having trouble with post-processing, try to figure out if you can replace the post-processing you're using with machine learning.\n\n 2. Some post-processing is applied separately to SJC and others, and some post-processing is applied only to SJC.\n\n 3. It is very noisy and we have not been able to use it well yet.\n\n 4. We have been working hard since we participated in the competition, but we haven't had a positive impact on LB yet. In other words, our LB were achieved using only post-processing and no domain knowledge at all. This might encourage people who don't have domain knowledge:)\n\n**Finally, let's enjoy this competition with all our strength for the remaining one month!**", "votes": 34 }, { "id": 1377737, "postDate": "2021-07-06T04:51:35.097Z", "content": "Thank you for hosting the important discussion!
\nHere are our team’s answers.
Are you taking advantage of machine learning?
\nYes we are. We are integrating learning-based algorithms as well as the rule-based ones into our estimation pipeline.
Do you have SJC-specific post-processing?
\nYes we have. We believe that one of the important parts of this competition is to optimize post process to each regions.
Are you using IMU data?
\nYes we are. We are also interested in using IMU to predict relative poses which was recently introduced here (https://www.kaggle.com/alvinai9603/predict-next-point-with-the-imu-data).
Have you succeeded in estimating the absolute position using RTKLIB etc.?
\nNot RTKLIB, but we are currently using our own baseline instead of the given one. The improvement was mainly achieved by tuning optimization functions and incorporating some domain knowledges about GNSS.
I spent a month learning GNSS. And there were a few unclear points in *_derived.csv files, I tried to generate equivalent files using raw GNSS logs and external data. I completed my first positioning program, but it didn't outperform the baseline provided by the host. (It performed equally well for some data, but far worse for other.) I think using the baseline is reasonable for GNSS non-expert, but I will try it again at the end of this competition.
\nSince last week, I've been focusing on post-processing and developed an optimization-based smoothing algorithm
\n(like this notebook, but more complicated). Currently my algorithm is almost equivalent to standard Kalman filter (smoother), but it might be useful for integrating various information in the future.
This competition is quite different than others on Kaggle - I've learned more about GPS in the last month than I ever thought I would :)
\nAre you taking advantage of machine learning?
\nNot yet - I've tried, but so far it hasn't been more effective than other methods. Also, most of my effort for the last month has been learning GPS trying to improve (or even just match!) the baseline
Do you have SJC-specific post-processing?
\nI've definitely looked at SJC separately as it is by far the hardest environment, but then I've folded those efforts back into my normal post-processing. That area has definitely affected certain thresholds that I'm using though.
Are you using IMU data?
\nNot yet - I've tried a couple of times and think I have a couple of good ideas, but the IMU data is tricky - I think it can be combined with delta gps data (kind of like the indoor competition, but different), but so far I haven't had much luck there
Have you succeeded in estimating the absolute position using RTKLIB etc.?
\nI've finally (FINALLY!) replicated the baseline (~7m error), but not using RTKLIB - I'm manually parsing the RINEX files and mixing that with the _derived information, and (after a lot of reading and errors and adjusting) have finally gotten similar results to the baseline.
I've tried to improve the baseline with some well known (to the GPS community) methods, but so far haven't succeed there - after learning so much about GPS this past month, I'm hopeful there though!
", "rawMarkdown": "This competition is quite different than others on Kaggle - I've learned more about GPS in the last month than I ever thought I would :)\n\n1. Are you taking advantage of machine learning?\nNot yet - I've tried, but so far it hasn't been more effective than other methods. Also, most of my effort for the last month has been learning GPS trying to improve (or even just match!) the baseline\n\n2. Do you have SJC-specific post-processing?\nI've definitely looked at SJC separately as it is by far the hardest environment, but then I've folded those efforts back into my normal post-processing. That area has definitely affected certain thresholds that I'm using though.\n\n3. Are you using IMU data?\nNot yet - I've tried a couple of times and think I have a couple of good ideas, but the IMU data is tricky - I think it can be combined with delta gps data (kind of like the indoor competition, but different), but so far I haven't had much luck there\n\n4. Have you succeeded in estimating the absolute position using RTKLIB etc.?\nI've finally (FINALLY!) replicated the baseline (~7m error), but not using RTKLIB - I'm manually parsing the RINEX files and mixing that with the _derived information, and (after a lot of reading and errors and adjusting) have finally gotten similar results to the baseline.\n\nI've tried to improve the baseline with some well known (to the GPS community) methods, but so far haven't succeed there - after learning so much about GPS this past month, I'm hopeful there though!", "votes": 8 }, { "id": 1383995, "postDate": "2021-07-11T11:34:00.203Z", "content": "Hi algorithmers :)
\nAre you taking advantage of machine learning?
\nYes, ML gave me a little more accurate result than baseline.
\nBut I use baseline location as an input feature which is similar to the notebook below, so my ML may be post processing of baseline rather than absolute position prediction model.
\nhttps://www.kaggle.com/jeongyoonlee/google-smartphone-decimeter-eda-keras-tpu
\nhttps://www.kaggle.com/asobod11138/gsdc-neuralnet-keras
Do you have SJC-specific post-processing?
\nYes. As described in below notebook and discussion, the quality of gnss data is quite different in each region.
\nhttps://www.kaggle.com/c/google-smartphone-decimeter-challenge/discussion/245160
\nhttps://www.kaggle.com/tyonemoto/data-divided-into-3-areas-downtown-highway-tree
Are you using IMU data?
\nYes, IMU data is useful to estimate relative / delta position, which were key factor of previous indoor location competition. I believe the same approach can be applied to this competition, so I’m keep working on establish IMU model.
Have you succeeded in estimating the absolute position using RTKLIB etc.?
\nNo, I never used RTKLIB.
Are you taking advantage of machine learning?
\nNo
Do you have SJC-specific post-processing?
\nNo. I tried but it was not successful
Are you using IMU data?
\nNo
Have you succeeded in estimating the absolute position using RTKLIB etc.?
\nNo, I tried but it was not successful
Thanks for sharing!!
\nI haven't spent that much time on this competition yet, but I hope my information is helpful to you all.
Are you taking advantage of machine learning?
\n-> Yes. I’m using it to process baseline data.
Do you have SJC-specific post-processing?
\n-> Yes. I have different post-process for different locations.
Are you using IMU data?
\n-> No.
Have you succeeded in estimating the absolute position using RTKLIB etc.?
\n-> No.
I was waiting for this topic to be posted😆
\n1.Are you taking advantage of machine learning?
\n Yes, I am using two ML models for post-processing.
\n (The two models look very similar and may be practically identical.)
2.Do you have SJC-specific post-processing?
\n Yes, I am focusing on this area and treating it differently from other areas in general.
\n (I call this area the \"Bermuda Triangle\".)
3.Are you using IMU data?
\n Yes, I use IMU for features in ML and other calculations.
4.Have you succeeded in estimating the absolute position using RTKLIB etc.?
\n I don't use RTKLIB,
\n but it has been partially successful in terms of improving baseline.
\n I believe there is still room for improvement in this.
Let's work hard for one more month!
", "rawMarkdown": "I was waiting for this topic to be posted😆\n\n1.Are you taking advantage of machine learning?\n Yes, I am using two ML models for post-processing.\n (The two models look very similar and may be practically identical.)\n\n2.Do you have SJC-specific post-processing?\n Yes, I am focusing on this area and treating it differently from other areas in general.\n (I call this area the \"Bermuda Triangle\".)\n\n3.Are you using IMU data?\n Yes, I use IMU for features in ML and other calculations.\n\n4.Have you succeeded in estimating the absolute position using RTKLIB etc.?\n I don't use RTKLIB, \n but it has been partially successful in terms of improving baseline.\n I believe there is still room for improvement in this.\n\nLet's work hard for one more month!", "votes": 5, "replies": [ { "id": 1376814, "postDate": "2021-07-05T11:06:38.503Z", "content": "Thanks for your comment. “Bermuda Triangle” is very funny. I'll call it that next time! lol
", "rawMarkdown": "Thanks for your comment. “Bermuda Triangle” is very funny. I'll call it that next time! lol", "votes": 3 } ] }, { "id": 1386167, "postDate": "2021-07-13T09:59:24.217Z", "content": "1.Are you taking advantage of machine learning?
\nNo.I tried, but it did not improve.
2.Do you have SJC-specific post-processing?
\nYes.I'm doing it like this discussion
3.Are you using IMU data?
\nYes.I use it to handle outliers.
4.Have you succeeded in estimating the absolute position using RTKLIB etc.?
\nNo.I didn't try.
Are you taking advantage of machine learning?
\nYes
Do you have SJC-specific post-processing?
\nNope but will try for sure
Are you using IMU data?
\nYes, sometimes
Have you succeeded in estimating the absolute position using RTKLIB etc.?
\nNot yet!
Q: Are you taking advantage of machine learning?
\nA: No
\nQ: Do you have SJC-specific post-processing?
\nA: Yes
\nQ: Are you using IMU data?
\nA: No
\nQ: Have you succeeded in estimating the absolute position using RTKLIB etc.?
\nA: Havent even tried
Thank you very much for your topic and sharing!
\nAre you taking advantage of machine learning?
\nYes, I use ML for post-processing.
Do you have SJC-specific post-processing?
\nSome post-processing is applied only to SJC.
Are you using IMU data?
\nYes.
Have you succeeded in estimating the absolute position using RTKLIB etc.?
\nNot yet.
Thanks @ryotayoshinobu, this thread is very useful!
\nMy responses below to the questions.
Are you taking advantage of machine learning?
\n-> Yes. Using an ensemble model.
Do you have SJC-specific post-processing?
\n-> No, do not have different post process for different location,
Are you using IMU data?
\n-> Yes. For handling outliers in the data.
Have you succeeded in estimating the absolute position using RTKLIB etc.?
\n-> Not yet, have to try that out.
Thank you for hosting the important discussion!
\nHere are our team’s answers.
Are you taking advantage of machine learning?
\nYes we are. We are integrating learning-based algorithms as well as the rule-based ones into our estimation pipeline.
Do you have SJC-specific post-processing?
\nYes we have. We believe that one of the important parts of this competition is to optimize post process to each regions.
Are you using IMU data?
\nYes we are. We are also interested in using IMU to predict relative poses which was recently introduced here (https://www.kaggle.com/alvinai9603/predict-next-point-with-the-imu-data).
Have you succeeded in estimating the absolute position using RTKLIB etc.?
\nNot RTKLIB, but we are currently using our own baseline instead of the given one. The improvement was mainly achieved by tuning optimization functions and incorporating some domain knowledges about GNSS.
I spent a month learning GNSS. And there were a few unclear points in *_derived.csv files, I tried to generate equivalent files using raw GNSS logs and external data. I completed my first positioning program, but it didn't outperform the baseline provided by the host. (It performed equally well for some data, but far worse for other.) I think using the baseline is reasonable for GNSS non-expert, but I will try it again at the end of this competition.
\nSince last week, I've been focusing on post-processing and developed an optimization-based smoothing algorithm
\n(like this notebook, but more complicated). Currently my algorithm is almost equivalent to standard Kalman filter (smoother), but it might be useful for integrating various information in the future.
This competition is quite different than others on Kaggle - I've learned more about GPS in the last month than I ever thought I would :)
\nAre you taking advantage of machine learning?
\nNot yet - I've tried, but so far it hasn't been more effective than other methods. Also, most of my effort for the last month has been learning GPS trying to improve (or even just match!) the baseline
Do you have SJC-specific post-processing?
\nI've definitely looked at SJC separately as it is by far the hardest environment, but then I've folded those efforts back into my normal post-processing. That area has definitely affected certain thresholds that I'm using though.
Are you using IMU data?
\nNot yet - I've tried a couple of times and think I have a couple of good ideas, but the IMU data is tricky - I think it can be combined with delta gps data (kind of like the indoor competition, but different), but so far I haven't had much luck there
Have you succeeded in estimating the absolute position using RTKLIB etc.?
\nI've finally (FINALLY!) replicated the baseline (~7m error), but not using RTKLIB - I'm manually parsing the RINEX files and mixing that with the _derived information, and (after a lot of reading and errors and adjusting) have finally gotten similar results to the baseline.
I've tried to improve the baseline with some well known (to the GPS community) methods, but so far haven't succeed there - after learning so much about GPS this past month, I'm hopeful there though!
", "votes": 8, "replies": [] }, { "id": 1383995, "author_name": "cocoinit23", "author_url": "", "post_date": "2021-07-11T11:34:00.203000", "content": "Hi algorithmers :)
\nAre you taking advantage of machine learning?
\nYes, ML gave me a little more accurate result than baseline.
\nBut I use baseline location as an input feature which is similar to the notebook below, so my ML may be post processing of baseline rather than absolute position prediction model.
\nhttps://www.kaggle.com/jeongyoonlee/google-smartphone-decimeter-eda-keras-tpu
\nhttps://www.kaggle.com/asobod11138/gsdc-neuralnet-keras
Do you have SJC-specific post-processing?
\nYes. As described in below notebook and discussion, the quality of gnss data is quite different in each region.
\nhttps://www.kaggle.com/c/google-smartphone-decimeter-challenge/discussion/245160
\nhttps://www.kaggle.com/tyonemoto/data-divided-into-3-areas-downtown-highway-tree
Are you using IMU data?
\nYes, IMU data is useful to estimate relative / delta position, which were key factor of previous indoor location competition. I believe the same approach can be applied to this competition, so I’m keep working on establish IMU model.
Have you succeeded in estimating the absolute position using RTKLIB etc.?
\nNo, I never used RTKLIB.
Are you taking advantage of machine learning?
\nNo
Do you have SJC-specific post-processing?
\nNo. I tried but it was not successful
Are you using IMU data?
\nNo
Have you succeeded in estimating the absolute position using RTKLIB etc.?
\nNo, I tried but it was not successful
Thanks for sharing!!
\nI haven't spent that much time on this competition yet, but I hope my information is helpful to you all.
Are you taking advantage of machine learning?
\n-> Yes. I’m using it to process baseline data.
Do you have SJC-specific post-processing?
\n-> Yes. I have different post-process for different locations.
Are you using IMU data?
\n-> No.
Have you succeeded in estimating the absolute position using RTKLIB etc.?
\n-> No.
I was waiting for this topic to be posted😆
\n1.Are you taking advantage of machine learning?
\n Yes, I am using two ML models for post-processing.
\n (The two models look very similar and may be practically identical.)
2.Do you have SJC-specific post-processing?
\n Yes, I am focusing on this area and treating it differently from other areas in general.
\n (I call this area the \"Bermuda Triangle\".)
3.Are you using IMU data?
\n Yes, I use IMU for features in ML and other calculations.
4.Have you succeeded in estimating the absolute position using RTKLIB etc.?
\n I don't use RTKLIB,
\n but it has been partially successful in terms of improving baseline.
\n I believe there is still room for improvement in this.
Let's work hard for one more month!
", "votes": 5, "replies": [ { "id": 1376814, "author_name": "penguin46", "author_url": "", "post_date": "2021-07-05T11:06:38.503000", "content": "Thanks for your comment. “Bermuda Triangle” is very funny. I'll call it that next time! lol
", "votes": 3, "replies": [] } ] }, { "id": 1386167, "author_name": "KenS", "author_url": "", "post_date": "2021-07-13T09:59:24.217000", "content": "1.Are you taking advantage of machine learning?
\nNo.I tried, but it did not improve.
2.Do you have SJC-specific post-processing?
\nYes.I'm doing it like this discussion
3.Are you using IMU data?
\nYes.I use it to handle outliers.
4.Have you succeeded in estimating the absolute position using RTKLIB etc.?
\nNo.I didn't try.
Are you taking advantage of machine learning?
\nYes
Do you have SJC-specific post-processing?
\nNope but will try for sure
Are you using IMU data?
\nYes, sometimes
Have you succeeded in estimating the absolute position using RTKLIB etc.?
\nNot yet!
Q: Are you taking advantage of machine learning?
\nA: No
\nQ: Do you have SJC-specific post-processing?
\nA: Yes
\nQ: Are you using IMU data?
\nA: No
\nQ: Have you succeeded in estimating the absolute position using RTKLIB etc.?
\nA: Havent even tried
Thank you very much for your topic and sharing!
\nAre you taking advantage of machine learning?
\nYes, I use ML for post-processing.
Do you have SJC-specific post-processing?
\nSome post-processing is applied only to SJC.
Are you using IMU data?
\nYes.
Have you succeeded in estimating the absolute position using RTKLIB etc.?
\nNot yet.
Thanks @ryotayoshinobu, this thread is very useful!
\nMy responses below to the questions.
Are you taking advantage of machine learning?
\n-> Yes. Using an ensemble model.
Do you have SJC-specific post-processing?
\n-> No, do not have different post process for different location,
Are you using IMU data?
\n-> Yes. For handling outliers in the data.
Have you succeeded in estimating the absolute position using RTKLIB etc.?
\n-> Not yet, have to try that out.