rotation error is large?

[Lzc6996]

Hi, i am trying to reproduce your results with the pretrained model you provided. And I find that, the posenet has very good translation error but large error in rotation. Is there something wrong?

[yzcjtr]

Hi @Lzc6996 , can you provide the translation error, rotation error and ATE error of your reproduced result on KITTI? Though I think translation error & rotation error are two different things and not directly comparable, ATE error or plotted trajectory may be better to discover the problem.

[Lzc6996]

@yzcjtr , thanks for your reply. I get ATE error about 0.012 in kitti seq.09 , just same as your report in GeoNet. But when I plot the trajectory, I find the trajectory is very bad. And then I replace the output rotation with gt rotation, I get a perfect trajectory. Did you plot the trajectory output by GeoNet? I just want to know if I did something wrong.

[yzcjtr]

Good observation. Did you perform scale matching between predictions and groundtruth when plotting the trajectory? If you perform such operation within each snippet of length 5, I believe you could get reasonable visualization. Can you post the plotted trajectory here?

[Lzc6996]

Yes, I did the scale matching.

[yzcjtr]

Nice, seems like the result is reasonable though not perfect. The main issue is accumulating error for VO algorithms. From the figure you can see the accumulated rotation error obviously. While I think you cannot tell accumulated translation error here for you have performed scale matching.

[Lzc6996]

In the slam problem, for example orb slam, usually it behaves better in rotation error rather than translation. Why can your posenet get a good result of translation without good rotation?

[Lzc6996]

And we are preparing a paper which wants to quote results of geonet . Do you think the results is fair? or should we do something to get a better results in rotation?

[yzcjtr]

Considering current clues you provide, I'm afraid we cannot conclude that "we get a good result of translation without good rotation", for:

1. Our PoseNet is better suited for a 5-frame snippet setting; Concatenation of predictions cannot avoid accumulating error issue; You can compare under 5 frame setting;
2. I would suggest you get quantitative results of "rotation error" or "translation error", like this paper does (and compare with orb-slam). That would be more precise & persuasive.

[Lzc6996]

Thank you for your patience.

[zacr0]

Yes, I did the scale matching.

Excuse me, can you share how did you plot both the predicted and GT trajectories? I've been trying to do it but my results are pretty inacurate.

[BernardoTeixeira]

Can any of you please share the code of how to plot the full trajectory from the inference output?

[DaddyWesker]

Same question as Bernardo's.

[Lzc6996]

@DaddyWesker @BernardoTeixeira
KITTI's odometry development kit can help u to do this work.