About the result of Replica dataset

[lengfeng-bot]

Hello, first thanks for the evaluation code, I ran the Replica dataset and the TUM dataset with your evaluation code, but something seems to be wrong with the results. The rmse of the APE w.r.t. rotation part obtained by all replica datasets is very large.The other results are normal, except for this one The results of the TUM dataset are normal. I wonder if there is some problem in the data format conversion. In the image below, the R of the third column of all the replica dataset reaches 2.8. R behaves normally on all TUM datasets The following is a metrics_traj in a scenario of the replica dataset,It can be seen that the error of the rotating part is abnormal

Can you tell me how to fix this, thank you very much!

[ZhangHX-2514]

Hi, we ran the evaluation code recently. And we got similar results, as shown in the picture below. What do you think about this issue now? Can you give us some advice?

[lengfeng-bot]

hello,recently I check the problem.First,I looked at the visualizations of the tracks,The reference track is the truth value of replica converted into tum data format,The other trajectory is the result of running it with Photo-slam.It is obvious that in xyz, there is no problem. but the degree of roll is reversed, the degree of pitch have some differences, but the error is small.so i think the main reason is the reversed roll angle And then I checked shapeReplicaGT.py . I don't quite understand these two sentences c2w[:3, 1] *= -1 c2w[:3, 2] *= -1 So I annotated these two sentences and ran shapeReplicaGT.py again. Then the rotation becomes normal we can see that the angles of the roll basically coincide,it's good. Finally, the APE of the rotating part is normal

[ZhangHX-2514]

Thank you very much for your reply. It's really helpful to us.

[ZhangHX-2514]

Hi, I noticed that before the code was modified, sse was very large, and after the modification, we all got much smaller sse values. However, when I evaluated the Monocular results in the same way, I got much larger sse values (shown below). So I wanna ask, does this mean that the Monocular evaluation results we obtained are still inaccurate? Thanks a lot!