Our team recorded public LB4.536 with only post processing based on baseline locations. However, we feel that there is a limit to post processing alone, and are interested in what approaches other participants are using.
\nSo, this is a discussion I would like to hear how everyone is approaching this competition.
\nIf you can answer any of the questions below, feel free to do so!
Question:
\nDo you use machine learning?
Do you use IMU data?
Are the absolute positions you are using them of the baseline?
In the Indoor Location & Navigation competition, a method called cost minimization was used to improve the accuracy of absolute positions using relative positions (see this page for more details). Therefore, do you use cost minimization?
Do you use RTKLIB(details)?
Here is our team's response:
\nDo you use machine learning?
\n-> We use it indirectly as part of the post-processing, but not directly to estimate the positions.
Do you use IMU data?
\n-> No, we don't use it now. We're working on the location estimation model.
Are the absolute positions you are using them of the baseline?
\n-> Yes. We are working on a model to estimate the positions from GNSS derived data, but currently baseline positions are more accurate.
Do you use cost minimization?
\n-> No. This is because we have not been able to estimate the relative positions.
Do you use RTKLIB?
\n-> No, but we are trying to use RTKLIB for RTK positioning
If you have any other questions, please leave them in the comment :)
\nOur team recorded public LB4.536 with only post processing based on baseline locations. However, we feel that there is a limit to post processing alone, and are interested in what approaches other participants are using.
\nSo, this is a discussion I would like to hear how everyone is approaching this competition.
\nIf you can answer any of the questions below, feel free to do so!
Question:
\nDo you use machine learning?
Do you use IMU data?
Are the absolute positions you are using them of the baseline?
In the Indoor Location & Navigation competition, a method called cost minimization was used to improve the accuracy of absolute positions using relative positions (see this page for more details). Therefore, do you use cost minimization?
Do you use RTKLIB(details)?
Here is our team's response:
\nDo you use machine learning?
\n-> We use it indirectly as part of the post-processing, but not directly to estimate the positions.
Do you use IMU data?
\n-> No, we don't use it now. We're working on the location estimation model.
Are the absolute positions you are using them of the baseline?
\n-> Yes. We are working on a model to estimate the positions from GNSS derived data, but currently baseline positions are more accurate.
Do you use cost minimization?
\n-> No. This is because we have not been able to estimate the relative positions.
Do you use RTKLIB?
\n-> No, but we are trying to use RTKLIB for RTK positioning
If you have any other questions, please leave them in the comment :)
\nGreat post, it's nice to see people are having the same trouble with the data as me 😂
\nMy approach so far sounds quite similar to you: just post-processing the given baseline positions. Also like you, I think I've reached about the limit that my approach can have right now.
\nI've tried a couple of times to get the gnss data to give better results, but since I'm new to gnss, it's quite complicated and difficult to get working right! I'm sure the final teams in 2 months will get gnss working correctly though and have much better accuracy than I do know, so I'll keep trying :)
\nNo machine learning yet - I tried, but couldn't get it to work as well as other post-processing techniques
No IMU data yet - again, I tried, but couldn't get it to work very well. I do think there is some signal there, but haven't be able to extract it yet. In the indoor competition, the phone was held by a person, so \"steps\" were visible on the acc data, which made it much easier to get a distance measurement from that. Since the phones are fixed to the car in this competition, I'm not sure such a calculation will be possible. Not sure yet.
Yes, I'm using the baseline positions that were given. Again, I tried to get a better baseline, but have failed so far 😂
Cost minimization: not yet; it's not worth it unless I can get IMU data to work better
RTKLIB: nope
Thank you for starting this discussion.
\nDo you use machine learning?
\n-> Yes. Today I have sent the first submission which uses ML model and it gave some improvement over processed baseline (LB Score: 5.430 -> 5.259). Model predicts relative positions using baseline and IMU data.
Do you use IMU data?
\n-> Yes. But gain from IMU data isn't very big so far. Baseline is much more important in my model.
Are the absolute positions you are using them of the baseline?
\n-> Yes
Do you use cost minimization?
\n-> No. I use Kalman Smoothing to combine relative positions with the baseline. But according to this notebook Kalman Smoothing and cost minimization are equivalent.
Do you use RTKLIB?
\n-> No
Thank you for referring to my notebook.
\nI'm trying to create a model that predicts locations from raw GNSS logs, but I'm still in the process of gathering information.
", "rawMarkdown": "Thank you for referring to my notebook.\n\nI'm trying to create a model that predicts locations from raw GNSS logs, but I'm still in the process of gathering information.", "votes": 10, "replies": [ { "id": 1341982, "postDate": "2021-06-09T06:04:24.927Z", "content": "Thanks for your comment and great notebook, @saitodevel01 !
\nI found the indoor competition very helpful.
\nI'm looking forward to competing with you in this competition again!
Thanks for sharing your great info!
\nI understood PostProcessing is very important in this competition.
I share my current state here too.
\nDo you use machine learning?
\n-> No, I tried some models and features, but not good.
Do you use IMU data?
\n-> No, I tried some models and features, but not good.
Are the absolute positions you are using them of the baseline?
\n-> Yes.
Do you use cost minimization?
\n-> Not yet., but I think this method is also useful in this competition if we calculate proper delta.
Do you use RTKLIB?
\n-> No.
LB 4.868 and 4th place at the time of writing a reply.
\nQ: Do you use machine learning?
\nA: Didnt have time for that yet except for a few experiments that costed me around ~30 subs, but I plan to return to it soon after solving some rl issues.
\nQ: Do you use IMU data?
\nA: Nope.
\nQ: Are the absolute positions you are using them of the baseline?
\nA: I have tried to use the GNSS logs, but files are so messy and sort of incomplete with different ways to merge them that I am waiting until somebody makes some of the GNSS usage public 😂
\nQ: Do you use cost minimization?
\nA: No, I suppose that I need to give a positive answer for the first question in order to be able to give a positive answer for this one.
\nQ: Do you use RTKLIB?
\nA: Not yet.
", "rawMarkdown": "LB 4.868 and 4th place at the time of writing a reply.\n\nQ: Do you use machine learning?\n\nA: Didnt have time for that yet except for a few experiments that costed me around ~30 subs, but I plan to return to it soon after solving some rl issues.\n\n\nQ: Do you use IMU data?\n\nA: Nope.\n\n\nQ: Are the absolute positions you are using them of the baseline?\n\nA: I have tried to use the GNSS logs, but files are so messy and sort of incomplete with different ways to merge them that I am waiting until somebody makes some of the GNSS usage public 😂\n\n\nQ: Do you use cost minimization?\n\nA: No, I suppose that I need to give a positive answer for the first question in order to be able to give a positive answer for this one.\n\nQ: Do you use RTKLIB?\n\nA: Not yet.\n", "votes": 7 }, { "id": 1340876, "postDate": "2021-06-08T10:08:00.960Z", "content": "Hi, thanks for sharing your insights!
\nI’m also trying to integrate the IMU information with GNSS using kalman filter. Since all the smartphones are fixed on a car, the motion model can be easily derived. I am currently using a state model as z = [x, y, yaw, vel, yawrate, acc].
\nBut I still cannot get reasonable results from the filter. Using optimization-based approach instead of filtering-based approach might be better.
\nBTW Im not using ML, RTK, or cost minimization. Im using the provided baseline. I tried to create my own baseline using gnss_derived files, but only could get up to 8.1.
", "rawMarkdown": "Hi, thanks for sharing your insights!\n\nI’m also trying to integrate the IMU information with GNSS using kalman filter. Since all the smartphones are fixed on a car, the motion model can be easily derived. I am currently using a state model as z = [x, y, yaw, vel, yawrate, acc]. \n\nBut I still cannot get reasonable results from the filter. Using optimization-based approach instead of filtering-based approach might be better.\n\nBTW Im not using ML, RTK, or cost minimization. Im using the provided baseline. I tried to create my own baseline using gnss_derived files, but only could get up to 8.1.", "votes": 8, "replies": [ { "id": 1340917, "postDate": "2021-06-08T10:33:46.027Z", "content": "Thanks for sharing your approach with us!
\nI\"ll reply on @colum2131 behalf.
Like you, We tried to estimate the azimuth using Kalman filter with IMU data.
\nHowever, the results were quite terrible.
\nWe are exploring better ways to use IMU data.
Using optimization-based approach instead of filtering-based approach might be better.
So We were wondering about the optimization based approach you mentioned.
\nDo you know of any articles that would be helpful?
We would appreciate it if you could let me know!
", "rawMarkdown": "Thanks for sharing your approach with us!\nI\"ll reply on [@colum2131](https://www.kaggle.com/columbia2131) behalf.\n\nLike you, We tried to estimate the azimuth using Kalman filter with IMU data.\nHowever, the results were quite terrible.\nWe are exploring better ways to use IMU data.\n\n`Using optimization-based approach instead of filtering-based approach might be better.`\n\nSo We were wondering about the optimization based approach you mentioned.\nDo you know of any articles that would be helpful?\n\nWe would appreciate it if you could let me know!", "votes": 1 }, { "id": 1340939, "postDate": "2021-06-08T10:46:28.887Z", "content": "\n\nLike you, We tried to estimate the azimuth using Kalman filter with IMU data
\n
Mostly the same as you…. It seems that noise filtering and bias calibration are critical if we try to use these sensor data.
\n\n\nSo We were wondering about the optimization based approach you mentioned.
\n
\n Do you know of any articles that would be helpful?
\n The method I meant above is something like the method used in the link you’ve posted (the cost-minimization one).
Unfortunately, I’m not sure whether there exists an article about an optimization approach in GNSS+IMU fusion.
\nI can give you an article about Vision+IMU fusion based on optimization instead: VINS-Mono
\nAlthough the article is not directly related to this competition (and also pretty complicated), it might be a bit of help to look into those equations.
It seems that noise filtering and bias calibration are critical if we try to use these sensor data.\nI agree with you that caliblation is important (manual calibration has improved prediction somewhat).
\nThe method I meant above is something like the method used in the link you’ve posted (the cost-minimization one).
I'm sorry. I mistakenly thought there were other approaches that combined optimization-based GNSS and IMU.
\nI can give you an article about Vision+IMU fusion based on optimization instead: VINS-Mono
And thanks for sharing the article.
\nI'll check out the article you shared :)
We still have a long way to go, but let's keep at it!
", "rawMarkdown": "```\nIt seems that noise filtering and bias calibration are critical if we try to use these sensor data.\n\n```\n\nI agree with you that caliblation is important (manual calibration has improved prediction somewhat).\n\n`The method I meant above is something like the method used in the link you’ve posted (the cost-minimization one).`\n\nI'm sorry. I mistakenly thought there were other approaches that combined optimization-based GNSS and IMU.\n\n`I can give you an article about Vision+IMU fusion based on optimization instead: VINS-Mono`\n\nAnd thanks for sharing the article.\nI'll check out the article you shared :)\n\nWe still have a long way to go, but let's keep at it!", "votes": 1 }, { "id": 1340983, "postDate": "2021-06-08T11:00:49.703Z", "content": "That's right!
\nGood luck to all of us :)
Hi Algorithms!! It ’s fun to compete with you again😄
\nI'm a beginner so it's not helpful,
\nDo you use machine learning?
\n-> No, absolute position estimation is at least not as good as baseline.
Do you use IMU data?
\n->Not used for Public LB at this time. I'm working now, i can roughly estimate, however it requires very high accuracy to contribute to the absolute position estimation of the host. I haven't reached that yet. When completed I would like to use it to cost minimization etc shape the route.
Are the absolute positions you are using them of the baseline?
\n-> yes
Do you use RTKLIB(details)?
\n-> no
Hi @columbia2131 . Thank you for a very useful discussion!
\nI was very surprised to hear that your team's score was achieved with only baseline based post-processing.
Let me answer your questions.
\nDo you use machine learning?
\n-> Yes. I'm using a model that predicts speed-related metrics for post-processing.
\n This method has worked well, but I believe there are smarter ways to do things.
Do you use IMU data?
\n-> No. I haven't tackled this yet, but I think it should be my next priority.
Are the absolute positions you are using them of the baseline?
\n-> Yes.
Do you use cost minimization?
\n-> No.
Do you use RTKLIB?
\n-> No.
Hi , @columbia2131
\nThanks for posting a great discussion.I also answer questions.
Hi @columbia2131 and congratulations to your team for the results so far.
\nSo, to answer your questions:
Did you used Kalman filter in your approach or process the baseline_locations data with another methods ?
", "rawMarkdown": "Hi @columbia2131 and congratulations to your team for the results so far.\nSo, to answer your questions:\n1. No, for now I have not used machine learning in this competition but I am considering for first data processing and then data inference\n2. I have tried to use the accelerometer data integrated into the Kalman filter but so far did not got good results.\n3. Yes\n4. Did not tried so far\n5. No, it is the first time I hear about RTKLIB but I will take a look.\n\nDid you used Kalman filter in your approach or process the baseline_locations data with another methods ?", "votes": 6, "replies": [ { "id": 1340859, "postDate": "2021-06-08T09:58:38.723Z", "content": "Thank you for your comment!
\n\n\nDid you used Kalman filter in your approach or process the baseline_locations data with another methods ?
\n
Our team uses the Kalman filter for the baseline location data after outlier processing, which is based on this page created by @dehokanta. And then, we do various other post processing.
\nI hope this helps.
Thanks for the great sharing.
\nMy score was low and I wasn't sure about it, so it took me a while to post.
Do you use machine learning?
\n-> No We are not using machine learning, we only use post processing.
Do you use IMU data?
\n-> No. We think IMU data is important and will include it in the future.
Are the absolute positions you are using them of the baseline?
\n-> Yes
Do you use cost minimization?
\n-> No
Do you use RTKLIB?
\n-> No
We try to approach each area differently, as shown in this notebook.
", "rawMarkdown": "Thanks for the great sharing.\nMy score was low and I wasn't sure about it, so it took me a while to post.\n\nDo you use machine learning?\n-> No We are not using machine learning, we only use post processing.\n\nDo you use IMU data?\n-> No. We think IMU data is important and will include it in the future.\n\nAre the absolute positions you are using them of the baseline?\n-> Yes\n\nDo you use cost minimization?\n-> No\n\nDo you use RTKLIB?\n-> No\n\nWe try to approach each area differently, as shown in[ this notebook.](https://www.kaggle.com/tyonemoto/data-divided-into-3-areas-downtown-highway-tree)\n\n\n\n", "votes": 3 }, { "id": 1348407, "postDate": "2021-06-14T03:23:47.137Z", "content": "Do you use machine learning? -> No
\nDo you use IMU data? -> Yes
\nAre the absolute positions you are using them of the baseline? -> No
\nDo you use cost minimization? No, extended Kalman filter to compute receiver position relative to nearby stations at known positions.
\nDo you use RTKLIB(details)? Yes and No. We're developing our own software similar to RTKlib ( ref: https://www.ion.org/gnss/abstracts.cfm?paperID=10489). We still have a lot of work to do, to optimize the solution. GNSS data from cellphones are a nightmare 😄. We haven't submitted our results as yet, but our locally computed score (~8) is worst than the filtered solution utilizing the baseline results.
\nIf you guys need support with RTKlib, I can try to provide some help.
", "rawMarkdown": "Do you use machine learning? -> No\n\nDo you use IMU data? -> Yes\n\nAre the absolute positions you are using them of the baseline? -> No\n\nDo you use cost minimization? No, extended Kalman filter to compute receiver position relative to nearby stations at known positions.\n\nDo you use RTKLIB(details)? Yes and No. We're developing our own software similar to RTKlib ( ref: https://www.ion.org/gnss/abstracts.cfm?paperID=10489). We still have a lot of work to do, to optimize the solution. GNSS data from cellphones are a nightmare 😄. We haven't submitted our results as yet, but our locally computed score (~8) is worst than the filtered solution utilizing the baseline results.\n\nIf you guys need support with RTKlib, I can try to provide some help.", "votes": 4 }, { "id": 1343430, "postDate": "2021-06-10T07:44:05.520Z", "content": "It is really a useful discussion. I have scored good, do support me in Kaggle.
\nDo you use machine learning?
\n-> Yes. I am into Machine Learning from long time but started submitting in contests recently.
Do you use IMU data?
\n-> Yes. But gain from IMU data isn't very big so far. Baseline is much more important in my model.
Are the absolute positions you are using them of the baseline?
\n-> Yes
Do you use cost minimization?
\n-> Yes sometimes.
Do you use RTKLIB?
\n-> No
Wow, That's very interesting.
", "rawMarkdown": "Wow, That's very interesting. ", "votes": -1 }, { "id": 1386431, "postDate": "2021-07-13T13:35:24.967Z", "content": "This is definitely one off Kaggle competition giving vibes of real world taste. Not everything can done with pure ML, but using it with subject matter expertise to use ML creatively!
", "rawMarkdown": "This is definitely one off Kaggle competition giving vibes of real world taste. Not everything can done with pure ML, but using it with subject matter expertise to use ML creatively! " }, { "id": 1378685, "postDate": "2021-07-06T17:49:39.827Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1376657, "postDate": "2021-07-05T08:54:05.330Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 1341058, "author_name": "chris", "author_url": "", "post_date": "2021-06-08T11:56:40.107000", "content": "Great post, it's nice to see people are having the same trouble with the data as me 😂
\nMy approach so far sounds quite similar to you: just post-processing the given baseline positions. Also like you, I think I've reached about the limit that my approach can have right now.
\nI've tried a couple of times to get the gnss data to give better results, but since I'm new to gnss, it's quite complicated and difficult to get working right! I'm sure the final teams in 2 months will get gnss working correctly though and have much better accuracy than I do know, so I'll keep trying :)
\nNo machine learning yet - I tried, but couldn't get it to work as well as other post-processing techniques
No IMU data yet - again, I tried, but couldn't get it to work very well. I do think there is some signal there, but haven't be able to extract it yet. In the indoor competition, the phone was held by a person, so \"steps\" were visible on the acc data, which made it much easier to get a distance measurement from that. Since the phones are fixed to the car in this competition, I'm not sure such a calculation will be possible. Not sure yet.
Yes, I'm using the baseline positions that were given. Again, I tried to get a better baseline, but have failed so far 😂
Cost minimization: not yet; it's not worth it unless I can get IMU data to work better
RTKLIB: nope
Thank you for starting this discussion.
\nDo you use machine learning?
\n-> Yes. Today I have sent the first submission which uses ML model and it gave some improvement over processed baseline (LB Score: 5.430 -> 5.259). Model predicts relative positions using baseline and IMU data.
Do you use IMU data?
\n-> Yes. But gain from IMU data isn't very big so far. Baseline is much more important in my model.
Are the absolute positions you are using them of the baseline?
\n-> Yes
Do you use cost minimization?
\n-> No. I use Kalman Smoothing to combine relative positions with the baseline. But according to this notebook Kalman Smoothing and cost minimization are equivalent.
Do you use RTKLIB?
\n-> No
Thank you for referring to my notebook.
\nI'm trying to create a model that predicts locations from raw GNSS logs, but I'm still in the process of gathering information.
", "votes": 10, "replies": [ { "id": 1341982, "author_name": "colum2131", "author_url": "", "post_date": "2021-06-09T06:04:24.927000", "content": "Thanks for your comment and great notebook, @saitodevel01 !
\nI found the indoor competition very helpful.
\nI'm looking forward to competing with you in this competition again!
Thanks for sharing your great info!
\nI understood PostProcessing is very important in this competition.
I share my current state here too.
\nDo you use machine learning?
\n-> No, I tried some models and features, but not good.
Do you use IMU data?
\n-> No, I tried some models and features, but not good.
Are the absolute positions you are using them of the baseline?
\n-> Yes.
Do you use cost minimization?
\n-> Not yet., but I think this method is also useful in this competition if we calculate proper delta.
Do you use RTKLIB?
\n-> No.
LB 4.868 and 4th place at the time of writing a reply.
\nQ: Do you use machine learning?
\nA: Didnt have time for that yet except for a few experiments that costed me around ~30 subs, but I plan to return to it soon after solving some rl issues.
\nQ: Do you use IMU data?
\nA: Nope.
\nQ: Are the absolute positions you are using them of the baseline?
\nA: I have tried to use the GNSS logs, but files are so messy and sort of incomplete with different ways to merge them that I am waiting until somebody makes some of the GNSS usage public 😂
\nQ: Do you use cost minimization?
\nA: No, I suppose that I need to give a positive answer for the first question in order to be able to give a positive answer for this one.
\nQ: Do you use RTKLIB?
\nA: Not yet.
", "votes": 7, "replies": [] }, { "id": 1340876, "author_name": "koji", "author_url": "", "post_date": "2021-06-08T10:08:00.960000", "content": "Hi, thanks for sharing your insights!
\nI’m also trying to integrate the IMU information with GNSS using kalman filter. Since all the smartphones are fixed on a car, the motion model can be easily derived. I am currently using a state model as z = [x, y, yaw, vel, yawrate, acc].
\nBut I still cannot get reasonable results from the filter. Using optimization-based approach instead of filtering-based approach might be better.
\nBTW Im not using ML, RTK, or cost minimization. Im using the provided baseline. I tried to create my own baseline using gnss_derived files, but only could get up to 8.1.
", "votes": 8, "replies": [ { "id": 1340917, "author_name": "213tubo", "author_url": "", "post_date": "2021-06-08T10:33:46.027000", "content": "Thanks for sharing your approach with us!
\nI\"ll reply on @colum2131 behalf.
Like you, We tried to estimate the azimuth using Kalman filter with IMU data.
\nHowever, the results were quite terrible.
\nWe are exploring better ways to use IMU data.
Using optimization-based approach instead of filtering-based approach might be better.
So We were wondering about the optimization based approach you mentioned.
\nDo you know of any articles that would be helpful?
We would appreciate it if you could let me know!
", "votes": 1, "replies": [] }, { "id": 1340939, "author_name": "koji", "author_url": "", "post_date": "2021-06-08T10:46:28.887000", "content": "\n\nLike you, We tried to estimate the azimuth using Kalman filter with IMU data
\n
Mostly the same as you…. It seems that noise filtering and bias calibration are critical if we try to use these sensor data.
\n\n\nSo We were wondering about the optimization based approach you mentioned.
\n
\n Do you know of any articles that would be helpful?
\n The method I meant above is something like the method used in the link you’ve posted (the cost-minimization one).
Unfortunately, I’m not sure whether there exists an article about an optimization approach in GNSS+IMU fusion.
\nI can give you an article about Vision+IMU fusion based on optimization instead: VINS-Mono
\nAlthough the article is not directly related to this competition (and also pretty complicated), it might be a bit of help to look into those equations.
It seems that noise filtering and bias calibration are critical if we try to use these sensor data.\nI agree with you that caliblation is important (manual calibration has improved prediction somewhat).
\nThe method I meant above is something like the method used in the link you’ve posted (the cost-minimization one).
I'm sorry. I mistakenly thought there were other approaches that combined optimization-based GNSS and IMU.
\nI can give you an article about Vision+IMU fusion based on optimization instead: VINS-Mono
And thanks for sharing the article.
\nI'll check out the article you shared :)
We still have a long way to go, but let's keep at it!
", "votes": 1, "replies": [] }, { "id": 1340983, "author_name": "koji", "author_url": "", "post_date": "2021-06-08T11:00:49.703000", "content": "That's right!
\nGood luck to all of us :)
Hi Algorithms!! It ’s fun to compete with you again😄
\nI'm a beginner so it's not helpful,
\nDo you use machine learning?
\n-> No, absolute position estimation is at least not as good as baseline.
Do you use IMU data?
\n->Not used for Public LB at this time. I'm working now, i can roughly estimate, however it requires very high accuracy to contribute to the absolute position estimation of the host. I haven't reached that yet. When completed I would like to use it to cost minimization etc shape the route.
Are the absolute positions you are using them of the baseline?
\n-> yes
Do you use RTKLIB(details)?
\n-> no
Hi @columbia2131 . Thank you for a very useful discussion!
\nI was very surprised to hear that your team's score was achieved with only baseline based post-processing.
Let me answer your questions.
\nDo you use machine learning?
\n-> Yes. I'm using a model that predicts speed-related metrics for post-processing.
\n This method has worked well, but I believe there are smarter ways to do things.
Do you use IMU data?
\n-> No. I haven't tackled this yet, but I think it should be my next priority.
Are the absolute positions you are using them of the baseline?
\n-> Yes.
Do you use cost minimization?
\n-> No.
Do you use RTKLIB?
\n-> No.
Hi , @columbia2131
\nThanks for posting a great discussion.I also answer questions.
Hi @columbia2131 and congratulations to your team for the results so far.
\nSo, to answer your questions:
Did you used Kalman filter in your approach or process the baseline_locations data with another methods ?
", "votes": 6, "replies": [ { "id": 1340859, "author_name": "colum2131", "author_url": "", "post_date": "2021-06-08T09:58:38.723000", "content": "Thank you for your comment!
\n\n\nDid you used Kalman filter in your approach or process the baseline_locations data with another methods ?
\n
Our team uses the Kalman filter for the baseline location data after outlier processing, which is based on this page created by @dehokanta. And then, we do various other post processing.
\nI hope this helps.
Thanks for the great sharing.
\nMy score was low and I wasn't sure about it, so it took me a while to post.
Do you use machine learning?
\n-> No We are not using machine learning, we only use post processing.
Do you use IMU data?
\n-> No. We think IMU data is important and will include it in the future.
Are the absolute positions you are using them of the baseline?
\n-> Yes
Do you use cost minimization?
\n-> No
Do you use RTKLIB?
\n-> No
We try to approach each area differently, as shown in this notebook.
", "votes": 3, "replies": [] }, { "id": 1348407, "author_name": "Garrett Seepersad", "author_url": "", "post_date": "2021-06-14T03:23:47.137000", "content": "Do you use machine learning? -> No
\nDo you use IMU data? -> Yes
\nAre the absolute positions you are using them of the baseline? -> No
\nDo you use cost minimization? No, extended Kalman filter to compute receiver position relative to nearby stations at known positions.
\nDo you use RTKLIB(details)? Yes and No. We're developing our own software similar to RTKlib ( ref: https://www.ion.org/gnss/abstracts.cfm?paperID=10489). We still have a lot of work to do, to optimize the solution. GNSS data from cellphones are a nightmare 😄. We haven't submitted our results as yet, but our locally computed score (~8) is worst than the filtered solution utilizing the baseline results.
\nIf you guys need support with RTKlib, I can try to provide some help.
", "votes": 4, "replies": [] }, { "id": 1343430, "author_name": "Arnab Dey", "author_url": "", "post_date": "2021-06-10T07:44:05.520000", "content": "It is really a useful discussion. I have scored good, do support me in Kaggle.
\nDo you use machine learning?
\n-> Yes. I am into Machine Learning from long time but started submitting in contests recently.
Do you use IMU data?
\n-> Yes. But gain from IMU data isn't very big so far. Baseline is much more important in my model.
Are the absolute positions you are using them of the baseline?
\n-> Yes
Do you use cost minimization?
\n-> Yes sometimes.
Do you use RTKLIB?
\n-> No
Wow, That's very interesting.
", "votes": -1, "replies": [] }, { "id": 1386431, "author_name": "Shahebaz Mohammad", "author_url": "", "post_date": "2021-07-13T13:35:24.967000", "content": "This is definitely one off Kaggle competition giving vibes of real world taste. Not everything can done with pure ML, but using it with subject matter expertise to use ML creatively!
", "votes": 0, "replies": [] }, { "id": 1378685, "author_name": "", "author_url": "", "post_date": "2021-07-06T17:49:39.827000", "content": "", "votes": 0, "replies": [] }, { "id": 1376657, "author_name": "", "author_url": "", "post_date": "2021-07-05T08:54:05.330000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1340647": "Our team recorded public LB4.536 with only post processing based on baseline locations. However, we feel that there is a limit to post processing alone, and are interested in what approaches other participants are using. \n\nSo, this is a discussion I would like to hear how everyone is approaching this competition.\nIf you can answer any of the questions below, feel free to do so!\n\n**Question:**\n\n1. Do you use machine learning?\n\n2. Do you use IMU data?\n\n3. Are the absolute positions you are using them of the baseline?\n\n4. In the [Indoor Location & Navigation competition](https://www.kaggle.com/c/indoor-location-navigation), a method called cost minimization was used to improve the accuracy of absolute positions using relative positions (see [this page](https://www.kaggle.com/saitodevel01/indoor-post-processing-by-cost-minimization) for more details). Therefore, do you use cost minimization?\n\n5. Do you use RTKLIB([details](http://www.rtklib.com/))?\n\n**Here is our team's response:**\n1. Do you use machine learning?\n-> We use it indirectly as part of the post-processing, but not directly to estimate the positions.\n\n2. Do you use IMU data?\n-> No, we don't use it now. We're working on the location estimation model.\n\n3. Are the absolute positions you are using them of the baseline?\n-> Yes. We are working on a model to estimate the positions from GNSS derived data, but currently baseline positions are more accurate.\n\n4. Do you use cost minimization?\n-> No. This is because we have not been able to estimate the relative positions.\n\n5. Do you use RTKLIB?\n-> No, but we are trying to use RTKLIB for RTK positioning\n\nIf you have any other questions, please leave them in the comment :)\n\n\n#### Reference\n* [indoor - Post-processing by Cost Minimization](https://www.kaggle.com/saitodevel01/indoor-post-processing-by-cost-minimizationA) created by [Akio Saito](https://www.kaggle.com/saitodevel01)\n", "1341058": "Great post, it's nice to see people are having the same trouble with the data as me 😂\n\nMy approach so far sounds quite similar to you: just post-processing the given baseline positions. Also like you, I think I've reached about the limit that my approach can have right now.\n\nI've tried a couple of times to get the gnss data to give better results, but since I'm new to gnss, it's quite complicated and difficult to get working right! I'm sure the final teams in 2 months will get gnss working correctly though and have much better accuracy than I do know, so I'll keep trying :)\n\n1. No machine learning yet - I tried, but couldn't get it to work as well as other post-processing techniques\n\n2. No IMU data yet - again, I tried, but couldn't get it to work very well. I do think there is some signal there, but haven't be able to extract it yet. In the indoor competition, the phone was held by a person, so \"steps\" were visible on the acc data, which made it much easier to get a distance measurement from that. Since the phones are fixed to the car in this competition, I'm not sure such a calculation will be possible. Not sure yet.\n\n3. Yes, I'm using the baseline positions that were given. Again, I tried to get a better baseline, but have failed so far 😂\n\n4. Cost minimization: not yet; it's not worth it unless I can get IMU data to work better\n\n5. RTKLIB: nope", "1341590": "Thank you for starting this discussion.\n\nDo you use machine learning?\n-> Yes. Today I have sent the first submission which uses ML model and it gave some improvement over processed baseline (LB Score: 5.430 -> 5.259). Model predicts relative positions using baseline and IMU data.\n\nDo you use IMU data?\n-> Yes. But gain from IMU data isn't very big so far. Baseline is much more important in my model.\n\nAre the absolute positions you are using them of the baseline?\n-> Yes\n\nDo you use cost minimization?\n-> No. I use Kalman Smoothing to combine relative positions with the baseline. But according to [this notebook](https://www.kaggle.com/saitodevel01/background-ideas-of-cost-minimization-notebook) Kalman Smoothing and cost minimization are equivalent.\n\nDo you use RTKLIB?\n-> No", "1341536": "Thank you for referring to my notebook.\n\nI'm trying to create a model that predicts locations from raw GNSS logs, but I'm still in the process of gathering information.", "1341168": "Thanks for sharing your great info!\nI understood PostProcessing is very important in this competition.\n\nI share my current state here too.\n\n1. Do you use machine learning?\n-> No, I tried some models and features, but not good.\n\n2. Do you use IMU data?\n-> No, I tried some models and features, but not good.\n\n3. Are the absolute positions you are using them of the baseline?\n-> Yes. \n\n4. Do you use cost minimization?\n-> Not yet., but I think this method is also useful in this competition if we calculate proper delta.\n\n5. Do you use RTKLIB?\n-> No.", "1340882": "LB 4.868 and 4th place at the time of writing a reply.\n\nQ: Do you use machine learning?\n\nA: Didnt have time for that yet except for a few experiments that costed me around ~30 subs, but I plan to return to it soon after solving some rl issues.\n\n\nQ: Do you use IMU data?\n\nA: Nope.\n\n\nQ: Are the absolute positions you are using them of the baseline?\n\nA: I have tried to use the GNSS logs, but files are so messy and sort of incomplete with different ways to merge them that I am waiting until somebody makes some of the GNSS usage public 😂\n\n\nQ: Do you use cost minimization?\n\nA: No, I suppose that I need to give a positive answer for the first question in order to be able to give a positive answer for this one.\n\nQ: Do you use RTKLIB?\n\nA: Not yet.\n", "1340876": "Hi, thanks for sharing your insights!\n\nI’m also trying to integrate the IMU information with GNSS using kalman filter. Since all the smartphones are fixed on a car, the motion model can be easily derived. I am currently using a state model as z = [x, y, yaw, vel, yawrate, acc]. \n\nBut I still cannot get reasonable results from the filter. Using optimization-based approach instead of filtering-based approach might be better.\n\nBTW Im not using ML, RTK, or cost minimization. Im using the provided baseline. I tried to create my own baseline using gnss_derived files, but only could get up to 8.1.", "1341136": "Hi Algorithms!! It ’s fun to compete with you again😄\n\nI'm a beginner so it's not helpful,\n\nDo you use machine learning?\n-> No, absolute position estimation is at least not as good as baseline.\n\nDo you use IMU data?\n->Not used for Public LB at this time. I'm working now, i can roughly estimate, however it requires very high accuracy to contribute to the absolute position estimation of the host. I haven't reached that yet. When completed I would like to use it to cost minimization etc shape the route.\n\nAre the absolute positions you are using them of the baseline?\n-> yes\n\nDo you use RTKLIB(details)?\n-> no", "1342289": "Hi @columbia2131 . Thank you for a very useful discussion!\nI was very surprised to hear that your team's score was achieved with only baseline based post-processing.\n\nLet me answer your questions.\n\n1. Do you use machine learning?\n-> Yes. I'm using a model that predicts speed-related metrics for post-processing.\n This method has worked well, but I believe there are smarter ways to do things.\n\n2. Do you use IMU data?\n-> No. I haven't tackled this yet, but I think it should be my next priority.\n\n3. Are the absolute positions you are using them of the baseline?\n-> Yes. \n\n4. Do you use cost minimization?\n-> No. \n\n5. Do you use RTKLIB?\n-> No.", "1341691": "Hi , @columbia2131 \nThanks for posting a great discussion.I also answer questions.\n\n1. Do you use machine learning?\nYes. my best score (LB:4.998) includes the machine learning solution.\nHowever, I don't have much confidence in it because I haven't had good results with CV. I plan to improve it in last 2 month.\n2. Do you use IMU data?\nNo.\n3. Are the absolute positions you are using them of the baseline?\nYes.\n4. Do you use cost minimization?\nNo. I tried, but it was difficult, so I gave up.\n5. Do you use RTKLIB?\nNo. I don't know about that. Thanks for sharing.", "1340817": "Hi @columbia2131 and congratulations to your team for the results so far.\nSo, to answer your questions:\n1. No, for now I have not used machine learning in this competition but I am considering for first data processing and then data inference\n2. I have tried to use the accelerometer data integrated into the Kalman filter but so far did not got good results.\n3. Yes\n4. Did not tried so far\n5. No, it is the first time I hear about RTKLIB but I will take a look.\n\nDid you used Kalman filter in your approach or process the baseline_locations data with another methods ?", "1365760": "Thanks for the great sharing.\nMy score was low and I wasn't sure about it, so it took me a while to post.\n\nDo you use machine learning?\n-> No We are not using machine learning, we only use post processing.\n\nDo you use IMU data?\n-> No. We think IMU data is important and will include it in the future.\n\nAre the absolute positions you are using them of the baseline?\n-> Yes\n\nDo you use cost minimization?\n-> No\n\nDo you use RTKLIB?\n-> No\n\nWe try to approach each area differently, as shown in[ this notebook.](https://www.kaggle.com/tyonemoto/data-divided-into-3-areas-downtown-highway-tree)\n\n\n\n", "1348407": "Do you use machine learning? -> No\n\nDo you use IMU data? -> Yes\n\nAre the absolute positions you are using them of the baseline? -> No\n\nDo you use cost minimization? No, extended Kalman filter to compute receiver position relative to nearby stations at known positions.\n\nDo you use RTKLIB(details)? Yes and No. We're developing our own software similar to RTKlib ( ref: https://www.ion.org/gnss/abstracts.cfm?paperID=10489). We still have a lot of work to do, to optimize the solution. GNSS data from cellphones are a nightmare 😄. We haven't submitted our results as yet, but our locally computed score (~8) is worst than the filtered solution utilizing the baseline results.\n\nIf you guys need support with RTKlib, I can try to provide some help.", "1343430": "It is really a useful discussion. I have scored good, do support me in Kaggle. \n\nDo you use machine learning?\n-> Yes. I am into Machine Learning from long time but started submitting in contests recently.\n\nDo you use IMU data?\n-> Yes. But gain from IMU data isn't very big so far. Baseline is much more important in my model.\n\nAre the absolute positions you are using them of the baseline?\n-> Yes\n\nDo you use cost minimization?\n-> Yes sometimes. \n\nDo you use RTKLIB?\n-> No", "1348172": "Wow, That's very interesting. ", "1386431": "This is definitely one off Kaggle competition giving vibes of real world taste. Not everything can done with pure ML, but using it with subject matter expertise to use ML creatively! ", "1378685": "", "1376657": "" } }