The code of another article about ai-imu-dr HOT 14 CLOSED

@Iqun1314 Hi, I am re-implementing RINS-W. Do you download the urban dataset and generate the ground-truth of RINS-W? I don't know how to generate the ground-truth of RINS-W by author. But according to my understanding of this article, the statistics I get are very different from the reported by the author. So, I hope to be able to communicate if you are re-implementing RINS-W.

from ai-imu-dr.

Ok,I will.
I had implemented another article https://arxiv.org/abs/1904.06064
The urban dataset is contained in the ZIP.
But I don't know how the author extract the dataset from the KITTI dataset.

from ai-imu-dr.

This is the paper of this repo. The code to extract the dataset is based on pyKitti.

Regarding the coe of RINS-W, we expect to put it online soonely :)

from ai-imu-dr.

This is the paper of this repo. The code to extract the dataset is based on pyKitti.

Regarding the coe of RINS-W, we expect to put it online soonely :)

Thank you for your reply. I'm looking forword to the code of RINS-W.

from ai-imu-dr.

@mbrossar Hi Martin, I am also very curious to know the details of the implementation of the RINS-W paper. The odometry performance is really good of RINS-W. Following up on the same ask since I saw your reply in March, do you plan to release the code soon in the near future? If so, roughly about when should we expect?

from ai-imu-dr.

I have put the code here, thank you for recalling me. There are few improvement with the paper, e.g. using CNN rather than LSTM, I let you see.

from ai-imu-dr.

@mbrossar Oh, thanks for such a prompt response! That was much quicker than I was expecting. Also thanks for uploading the code!
Just two quick questions:
In your paper it's using LSTM and in the code I see it's CNN. So you mean that the LSTM network has better performance than CNN?

I see you create the other repo yesterday/today. Seems that the code in that repo is not complete. In one file I see

from src.utils import pload, pdump, yload, ydump, mkdir, bmv
from src.utils import bmtm, bmtv, bmmt, pltt, plts, axat, pltt, plts
from src.lie_algebra import SO3, CPUSO3
from src.iekf import RecorderIEKF as IEKF

While lie_algebra.py file is there, I did not find iekf.py and utils.py files in the repo. Just to confirm with you.

from ai-imu-dr.

No, it is the opposite, I mean that CNN network has better performance than LSTM.
I am currently updated the second repo, you can now find iekf.py and utils.py files.

from ai-imu-dr.

@mbrossar now I can see all the files, thanks!

Maybe "performance" was not the right word I should use. In the README file, you said the CNN is faster to train than LSTM. I am curious if LSTM has better prediction accuracy than CNN. Could you comment on that from the validation test you did?

from ai-imu-dr.

I obtained better prediction accuracy with the CNN along with better time execution.

In the readme, I indicate CNN is faster to train than LSTM, which is in fact true. As performance of LSTM and CNN depends on hyperparameter, I can just say that, on my experiments, CNN were better.

from ai-imu-dr.

Can you confirm if Brossard's code on:

https://github.com/mbrossar/RINS-W

has the python tools by which to convert a dataset in KITTI format into Brossard's
particular python pickle format ready to be injected into his program in ai-imu-dr ?

from ai-imu-dr.

@Iqun1314 From your posts, I think you had access to the training data and the trained model.
Could you please share it with us?

Apparently, many followers of AI-IMU DR including the author Martin are looking for these data and model.
You may upload it to a cloud service provider like google drive or baidu cloud, and share the link with us.
Thank you so much.

from ai-imu-dr.

I found out that Brossard's code has the means to transform KITTI files into Brossard's pickle files which are composed of 5 python dictionaries apart from time. It is within main_kitty.py. Now his pickle file data dictionaries are made of u (accel+gyro), and quantities like p_gt which depend on lat/lon , v_gt is velocity Rot_gt are rotation matrices depending on roll, pitch and yaw. These are indexed and synchronized with the timestamp dictionary 't'. What I don't know is how GPS outage is modeled. Brossard's original data could tell us that. How do you work a test case where you only have u? ie. u and no GPS? For KITTI files, the standard is to replace the last 3 integer numbers with -1, but Brossard's code ignores such numbers. Can anyone tell me out GPS outage is modeled insofar as the test pickle files are concerned?

from ai-imu-dr.

Pickle files are now available at new url, please see the readme

from ai-imu-dr.