SKD: Unsupervised Keypoint Detection for Point Clouds using Saliency Estimation

In this work we present a novel keypoint detector that uses saliency to determine the best candidates from point clouds for tasks such as registration and reconstruction. The approach can be applied to any differentiable deep learning descriptor by using the gradients of that descriptor with respect to the input to estimate an initial set of candidate keypoints. By using a neural network over the set of candidates we can also learn to refine the point selection until the actual keypoints are obtained. The key intuition behind this approach is that keypoints need to be determined based on how the descriptor behaves given a task and not just because of the geometry that surrounds a point. To improve the performance of the learned keypoint descriptor we combine the saliency, the feature signal and geometric information from the point cloud to allow the network to select good keypoint candidates. The approach was evaluated on two large LIDAR datasets - the Oxford RobotCar dataset and the KITTI datasets, where we obtain up to 50% improvement over the state-of-the-art in both matchability score and repeatability. This results in a higher inlier ratio and a faster registration without compromising metric accuracy.