USIP: Unsupervised Stable Interest Point Detection From 3D Point Clouds

Abstract

In this paper, we propose the USIP detector: an Unsupervised Stable Interest Point detector that can detect highly repeatable and accurately localized keypoints from 3D point clouds under arbitrary transformations without the need for any ground truth training data. Our USIP detector consists of a feature proposal network that learns stable keypoints from input 3D point clouds and their respective transformed pairs from randomly generated transformations. We provide degeneracy analysis of our USIP detector and suggest solutions to prevent it. We encourage high repeatability and accurate localization of the keypoints with a probabilistic chamfer loss that minimizes the distances between the detected keypoints from the training point cloud pairs. Extensive experimental results of repeatability tests on several simulated and real-world 3D point cloud datasets from Lidar, RGB-D and CAD models show that our USIP detector significantly outperforms existing hand-crafted and deep learning-based 3D keypoint detectors. Our code is available at the project website. 1 * now at nuTonomy: an APTIV company.

Unsupervised Learning of Category-Specific Symmetric 3D Keypoints from Point Sets

“…More precisely, we want to learn the function Π C : S i → P i as a neural network of parameters θ, using the supervisory signal from Φ C . In regard to learning keypoints on pointsets, recent work [23] trains a Siamese network to predict order-agnostic keypoints stable to rotations for rigid objects [23].We use a similar network architecture as in [23] that is based on PointNet [40] but we do not use the Siamese training. The overview of the network is shown in Fig.…”

“…That being said, many keypoints are defined manually, while considering their semantic locations such as facial landmarks and human body joints, to serve and simplify the problem at hand. To further benefit from their widespread utility, several attempts have been made on learning to detect keypoints [16][17][18][19][20], as well as on automatically discovering them [21][22][23][24]. In this regard, the task of learning to detect keypoints from several supervision examples, has achieved many successes.…”

“…The challenges are exacerbated when the practical cases of misaligned data and unsupervised learning are considered. Related works do not address all these problems, but instead opt for; dropping category-specificity and using aligned data [23], employing manual supervision on 2D images [17], using aligned 3D and multiple 2D images with known pose [24]. The latter method achieves category-specificity without explicitly reasoning on the shapes.…”

“…2 shows the basic overview of our training strategy. Note that we do not require the Siamese-like architecture as in [4,23]. Using multiple categories from four benchmark datasets, we evaluate the quality of our learned keypoints both quantitatively and with qualitative visualization.…”

Low Overlapping Point Cloud Registration Using Line Features Detection

“…This explicitly introduces the general assumption of a point-cloud registration problem; that the size of an overlapping area is very large and only a minor correction in translation and rotation is sought [4]. Generally, the solution in a high overlapping point-cloud consists of keypoints detection [7][8][9], descriptors calculation [10][11][12] around each of the keypoints and running an Iterative Closest Point (ICP) algorithm [13,14] to find a transformation that pair-wise matches the individual descriptors. When the overlapping area is small, as in our case, it is difficult to reliably find the matching keypoints in the two-point-clouds, which is an essential step in almost all of the existing point-cloud registration approaches.…”