SKD: Keypoint Detection for Point Clouds Using Saliency Estimation

We present SKD, a novel keypoint detector that uses saliency to determine the best candidates from point cloud for tasks such as registration and reconstruction. The approach can be applied any differentiable deep learning descriptor by using gradients of with respect 3D position input points measure their saliency. is combined original context information in neural network, which trained learn robust candidates. key intuition behind this keypoints are not extracted solely result geometry surrounding point, but also take into account descriptor's response. was evaluated on two large LIDAR datasets - Oxford RobotCar dataset KITTI dataset, where we obtain up 50% improvement over state-of-the-art both matchability repeatability. When performing sparse matching computed our method achieve higher inlier ratio faster convergence.