Update README.md

Clarification on robust parameter estimation

README.md

@@ -5,7 +5,7 @@ Our ML4AD @ NeurIPS 2022 paper: [Stress-Testing Point Cloud Registration on Auto
 To train and test rigid point-cloud registration (PCR) algorithms with automotive LiDAR datasets, one must select a subset of point-cloud pairs. Previously, simple heuristics were used to create such registration sets, e.g., using a spacing of 10 meters. We provide a smart algorithm that selects challenging sets, that include a balanced sampling from the various situations that appear in a dataset (various offsets, rotations, etc.). 
 ![Screenshot from 2022-11-14 08-21-43](https://user-images.githubusercontent.com/12913832/201589994-249eefe2-2707-4e48-8e93-03f8abd7277b.png)
 
-We provide registration sets that were produced by our algorithm, for the Apollo-Southbay and NuScenes datasets. In our paper, we use these sets to train and benchmark some recent and popular registration algorithms. We use FCGF deep features with a variety of robust motion-estimation algorithms. Surprisingly, we find that the fastest and most accurate results come not from recent algorithms such as Teaser++ and PointDSC, but rather from a modern version of RANSAC. 
+We provide registration sets that were produced by our algorithm, for the Apollo-Southbay and NuScenes datasets. In our paper, we use these sets to train and benchmark some recent and popular registration algorithms. All algorithms use FCGF deep features, and differ in their method for robustly estimating the 6-DOF motion. Surprisingly, we find that the fastest and most accurate results come not from recent algorithms such as Teaser++ and PointDSC, but rather from a modern version of RANSAC. 
 
 ![time_and_recall_comparison_B_to_B_tight](https://user-images.githubusercontent.com/12913832/201589682-48c5cc9e-eb58-4e3a-9c01-058c58832b14.png)