Question: How are the 200 random point cloud pairs sampled for KITTI evaluation? #3
Description
Hi, thank you for the excellent work and open-source code!
I am using TCF as a baseline in my registration pipeline and trying to reproduce your KITTI evaluation results. In the paper (Section IV-A), it states:
"We randomly sample 200 pairs of point clouds from each sequence."
I noticed the sample data in the repo (data/KITTI/09/) contains pairs with frame gaps of 11, 12, and 19 (e.g., 1391-1380, 833-821, 210-191), which suggests the pairs are not simply consecutive frames.
Could you clarify the pair sampling protocol? Specifically:
- Is there a maximum distance or frame gap constraint between the two frames in a pair? (e.g., pairs within 10m ground-truth distance, or within N frames)
- Is there a minimum overlap ratio requirement to ensure the pair is registrable?
- Could you share the pair generation script or the list of (frame_i, frame_j) indices used in the paper's experiments?
This would help us ensure a fair comparison. Thank you!