Fig 1.
Data from the experimental protocol.
Point clouds from two datasets—(top) Gazebo with overlap 0.9 and (bottom) ETH with overlap 0.59. Reading and reference point clouds (left) prior registration and (right) aligned according to ground truth. The reference is displayed in blue, the reading in orange tones.
Fig 2.
(a) spherical support with local reference frame, (b) 8 orientation bins, (c) 4 spatial bins.
Fig 3.
Repeatability of keypoints from (top) points and (bottom) normals for each saliency measure.
Fig 4.
Average displacement eq of the corresponding local reference frames for (top) sign disambiguation methods and (bottom) pairs of disambiguated vectors ensuring right-handedness of the basis.
Fig 5.
Point cloud registration accuracy.
Distribution of (left) rotation and (right) translation errors for (top) all reading-reference pairs from hard poses and for (bottom) the pairs with overlap at least 0.75. A50, A75, and A95 denote the 50th, 75th, and 95th percentiles.
Table 1.
Quantile statistics of registration errors.
Fig 6.
Consensual feature correspondences.
The reading and reference point clouds are displayed in orange and blue tones, respectively, aligned with each other using the ground-truth pose. Black lines connect the corresponding features from the consensual set, i.e., the inliers, marked by red circles and blue squares. The markers would be concentric in case of a perfect match. (left) Accurate pose estimate from 19 inliers in Gazebo point clouds with overlap 0.5. (right) Inaccurate translation estimate from 99 inliers in ETH point clouds with overlap 0.67.
Table 2.
Average errors and running times of registration methods.