Question about the Paper

[yuvalH9]

Hi @fabiopoiesi , I really enjoyed reading your paper and thanks for supplying your code. After reading the paper I have some question which are still unclear to me. 1) In the bottleneck part you explain how you find potential matched points within the corresponding patches using alpha and alpha' that their corresponding feature cross some threshold. It is not clear to me when you use this idea in the training pipeline, or you just mentioned that for general analysis?

2) In Equ (6), it is not clear to me how you define the sets C_pos and C_neg and how you generate the feature vectors f. Does C_pos are a set of matching patches with their network output?

Thanks Yuval

[fabiopoiesi]

Hi @yuvalH9,

thanks for your interest in the paper.

1. As you mentioned, I just made these observations for sake of analysis and to motivate the thresholding steps done afterwards. I am not using alpha and alpha' in the actual computation of the descriptors.
2. f is the descriptors the network outputs. C_pos is a pair of descriptors created from two corresponding patches. These corresponding patches are extracted from two point clouds that have an overlap greater than 30%. C_neg are all the other descriptors within a batch (of course excluding C_pos).

I hope this helps.

[yuvalH9]

@fabiopoiesi thanks! it helps a lot indeed!

Another unclear issue, I'm still not sure where you are using p_rho in your training pipeline. You are ignoring patches with rho less than that threshold when you are calculating the Triplet loss? The pdf of rho is estimated from all the patches in a single batch?

Yuval

[fabiopoiesi]

Sorry for the late reply. I totally missed it. p_rho is used to filter out patches during registration, see Tab. V of the paper where different values are used.