zjuluolun / bevplace Goto Github PK

Hi Lunluo,

It seems that the number of the KITTI dataset in the Table 1 is different from the raw dataset downloaded from the official website. For example, there are 4541 bin files from 0000 to 4500 for the sequence 00. But in the Table 1, you mention that it's 0-3000 for the database, and 3200-4650 for the query. Have you made any changes to the dataset？

Thanks for your attention.

Hi, @zjuluolun .
Thank you for open-sourcing such excellent work.
I encountered some issues while reproducing the BEVPlace’s performance and would like to ask for some advice. I was able to reproduce the results very well on kitti 00, 02,05, and 06. However, when testing the performance of BEVPlace on sequences 07 and 08, I found a slight degradation in performance, as shown in the PR curve. I would like to know the reasons for this. Could it be an error in my reproduction code?

I would like to express my gratitude for sharing your research and data openly. However, I have encountered an issue that I believe warrants discussion. While conducting tests using the provided network weights on the dataset you shared, I observed that the achieved accuracy did not align with the precision levels stated in your paper. This discrepancy has prompted me to seek further clarification and insight into potential factors contributing to this divergence.
The detail results are as follow:
KITTI | 00 | 02 | 05 | 06 | Mean
paper | 99.7 | 98.1 | 99.3 | 100 | 99.275
my test | 98.53 | 95.16 | 98.88 | 100 | 98.1425

Hello, thank you for sharing your excellent work. I'm a little confused about the scale features in your work.
You have mentioned in your paper that "Our main modification to GIFT is to remove the scale features since there is no scale difference between BEV images". However, after I read your code, I don't quite understand how the scale features are removed. It seems that the scale features are still in the input.
Can you explain it? Thanks!

How do you get the ground truth of the loop. I'm interested in your work, in the process of reproduction, how do you get the ground truth of loops in datasets such as kitti? The number of true loops and the position of the two frames corresponding to the loop are important for the calculation of evaluation indicators such as accuracy and recall. Can you provide your method or code?

Hi zjuluolun,

Thanks for the code of your precious BEVPlace. I saw some application that can be done using this algorithm, such as Place Recognition, Loop closing detection, and pose estimator.

In terms of Place Recognition and Loop Closing detection this algorithm needs previous global map database, right? However, I'm curious about SLAM. If we don't have a map database yet, how can we apply this code to a SLAM (Mapping) problem?

Thank you

Hi Lunluo,

It seems the calculation of "Mean AR@1" and "Mean AR@1%" in Tab.5 is not correct.

(96.5 + 96.9 + 92.3 +95.3)/4 = 95.25
(99.0 + 99.7 + 98.7 + 99.5)/4 = 99.225

Thanks for your attention.

@zjuluolun Hello, I am exploring the BEVPlace project and find it very useful.
However, I noticed that the training scripts for the models are not provided in the repository. Would it be possible for you to share the training scripts used to train the models? This would be very helpful for me and other users who are interested in experimenting with the models. Thank you in advance for your time and consideration.