Questions on reproducing BEVPlace2 (negative threshold, ResNet-34 init, checkpoint choice, RANSAC/ICP)

[Dongyun-kang0]

Hello zjuluolun,

Thank you for the excellent paper and code. I am trying to reproduce the results and have a few clarifications:

[1.Negative threshold]
The paper states negatives are all candidates beyond 5 m, but the code appears to use 7 m. Which value is correct?

[2.Initialization for the cropped model]
Did you use ImageNet-pretrained weights for ResNet-34 when training the cropped model? If so, any layers frozen at the start?

[3.Reported metrics]
For the recall and success rate in the paper, which checkpoint produced those numbers? Were they from a single run or averaged across multiple seeds/runs?

[4.Pose estimation details]
Did you run RANSAC followed by SVD-ICP?
4.1) If yes, could you share the key settings: inlier threshold, max iterations, correspondence strategy, convergence criteria, and any downsampling or rejection steps?

Any pointers to exact config files or commit hashes would help.

Best regards,
Dongyun
KAIST, MSC Lab, Autonomous Driving Vision Team