Accurate Point Cloud Registration with Robust Optimal Transport

Authors: Zhengyang Shen, Jean Feydy, Peirong Liu, Ariel H Curiale, Ruben San Jose Estepar, Raul San Jose Estepar, Marc Niethammer

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our proposed models achieve SOTA results for scene flow on Kitti [78, 77] and for point-cloud-based lung registration on Dir Lab-COPDGene [19]. Notably, we show that Rob OT is highly suited to fine-tuning tasks: it consistently turns good matchings into nearly perfect registrations at an affordable numerical cost.
Researcher Affiliation Collaboration Zhengyang Shen UNC Chapel Hill zyshen@cs.unc.edu Jean Feydy Imperial College London jfeydy@ic.ac.uk Peirong Liu UNC Chapel Hill peirong@cs.unc.edu Ariel Hernán Curiale Harvard Medical School acuriale@bwh.harvard.edu Ruben San José Estépar Harvard Medical School rubensanjose@bwh.harvard.edu Raúl San José Estépar Harvard Medical School rjosest@bwh.harvard.edu Marc Niethammer UNC Chapel Hill mn@cs.unc.edu
Pseudocode No The paper describes algorithms and methods in detail using equations and textual explanations but does not provide structured pseudocode or algorithm blocks.
Open Source Code Yes Our code and dataset are available online at: https://github.com/uncbiag/robot.
Open Datasets Yes We train on the synthetic Flying3D dataset [75], which is made up of multiple moving objects that are sampled at random from Shape Net.
Dataset Splits Yes For all of our experiments, we randomly sample 600 training and 100 validation cases from this large collection of unannotated patients.
Hardware Specification Yes As detailed in Suppl. A.6, all run times were measured on a single GPU (24GB NVIDIA Quadro RTX 6000).
Software Dependencies Yes We then combine Eq. (3) with Eq. (4) to compute the Rob OT vectors vi and weights wi with O(N + M) memory footprint using the Ke Ops library [20, 37] for Py Torch [86] and Num Py [117].
Experiment Setup Yes We follow the same experimental setting as in [126, 49], with full details provided in Suppl. A.4.3: 1. We train on the synthetic Flying3D dataset [75], which is made up of multiple moving objects that are sampled at random from Shape Net. We take 19,640 pairs of point clouds for training, with dense ground truth correspondences. 2. We evaluate on 142 scene pairs from Kitti, a real-world dataset [78, 77]. We conduct experiments using 8,192 and 30k points per scan, sampled at random from the original data.