The result of ATE

[Lannist]

Hello, urbste. Thanks for your great contributions. I have tested your public dateset IndoorDynamic and found some problems.

When I use three cameras , the tracking is very robust, but the result of ATE is not well. The reuslt of IndoorDynamic in your paper(MULTICOL-SLAM - A MODULAR REAL-TIME MULTI-CAMERA SLAM SYSTEM) is 1.8cm , but the result I got is about 20cm. I've tried it many times.

Firstly, I did not change any parameters. I doubt that I am doing something wrong. Can you help me explain why I got the result so bad?

[urbste]

Hi. First, it is important to run the SLAM on a powerful machine (>i5) with enough threads. Results highly depend on that. Second, I know that my colleagues changed and fixed some stuff on the dataset. How did you evaluate the ATE? At the moment I don't have the means to check this myself, but as soon as can I will let you know.

[Lannist]

First, thanks for your timely reply. I tested the code on a powerful computer(i7) with enough threads. I think the hardware configuration is good enough. Second, I evaluate the ATE by the the public python script of TUM. The download link is ： http://vision.in.tum.de/data/datasets/rgbd-dataset/tools Did you use the same method to evaluate the ATE？

Now，I wonder if I have used the wrong parameters. Because the truth trajectory and the estimated trajectory are estimated to be very similar in shape. Scale problem or dataset？

I found there has other Scene‘s data（Indoor static, Outdoor static）in LafiDa Datasets. Can you give me the link page to download？

Thanks again！

[Lannist]

I think I found the problem. Your paper (LaFiDa—A Laserscanner Multi-Fisheye Camera Dataset) says that there always have a transformation matrix Mrbt ( in Sec4.4) which represents the transformation of the Rigid body to MCS.

But your public dataset(IndoorDynamic) don't contain this calibration file. So , I can't get the right result. The truth trajectory and the estimated trajectory are looks very similar in shape, but there is lack of a transformation matrix between them(Extrinsic Rigid Body to MCS Calibration).

Can you Uploaded the calibration file to your git? Looking forward to your reply.

Thanks again.

[jingdongZhang]

sorry,when ran this project,I met this problem.

System shutdown Requested to finish local mapper, loop closer and map publisher MakeCurrent: Not available with non-pangolin window. All threads stopped...

median tracking time: 0.0284605 mean tracking time: 0.0295164

Saving MKF trajectory to MKFTrajectory.txt ...

Saving MultiKeyframe trajectory and all MCS poses with timestamps

[squashking]

@jingdongZhang Hi Jiongdong, did you solve the issue? I ran into the same issue when running the example:

Start processing sequence ... Images in the sequence: 0

System shutdown Requested to finish local mapper, loop closer and map publisher All threads stopped...

[zc966]

@Lannist Hi Lannist, did you solve the issue? I have used the transformation matrix, but the result I got is still about 13cm.