Apple has won a Project Titan patent relating to techniques for implementing a 3D voxel feature detection network
Back in November 2017, Patently Apple posted a report titled "Apple Research Paper Touts VoxelNet as being Superior to LiDAR Regarding Autonomous Vehicle 3D Detection Methods." Apple's research paper noted that "Experiments on the KITTI car detection benchmark show that VoxelNet outperforms the state-of-the-art LiDAR based 3D detection methods by a large margin."
Today the U.S. Patent and Trademark Office officially granted Apple a patent titled "Voxel-based feature learning network" which relates to systems and algorithms for machine learning and more specifically to 3D object detection. More specifically, the patent covers solutions based on systems and algorithms that can accurately detect 3D objects in point clouds that may be considered a central problem in many real-world applications, such as autonomous navigation, housekeeping robots, and/or augmented/virtual reality.
Apple's invention covers methods, systems and/or techniques for implementing a 3D voxel feature learning/detection network that may be configured to learn voxel features automatically from raw point cloud data (e.g., a LiDAR point cloud). The 3D voxel feature learning/detection network may capture subtle 3D shape information that is typically lost due to quantization or due to other processing steps in previous systems.
In some embodiments, integrating a 3D voxel feature learning/detection network may lead to improved mean average object detection precision (mAP).
For example, mean average object detection precision (mAP) may be improved from 88.0%.fwdarw.89.2% and/or reduced mean orientation error may be reduced from 2.5 deg.fwdarw.0.82 deg as compared to previous point cloud object detection networks.
In addition, a 3D voxel feature learning/detection network may improve prediction stability through time. Additionally, a system configured to implement a 3D voxel feature learning/detection network, as described in Apple's granted patent, may be considered a generic feature learning module that may be configured to seamlessly integrate and train jointly with many other 3D machine learning models, such as: 3D object detection, 3D scene understanding, 3D point cloud matching, 3D object/human pose estimation, etc.
Accurately detecting 3D objects in point clouds may be considered a central problem in many real-world applications, such as autonomous navigation, housekeeping robots, and/or augmented/virtual reality.
Apple's patent FIG. 1A below is a high-level diagram illustrating end-to-end learning between a voxel feature learning/detection network and an object detection network; FIG. 1B is a high-level diagram illustrating contrasts between a voxel feature learning/detection network and other systems that utilize height quantization to generate a multi-channel image.
Apple's patent FIG. 10 above illustrates an example of training data for vehicle detection that may be used to train a voxel feature learning/detection network; FIG. 11 illustrates example types of objects that may be identified by an object detection system that includes a voxel feature learning/detection network, according to some embodiments.
Obviously, this patent isn't a consumer-friendly read. However, engineers and AI researchers may want to review Apple's granted 10,970,518 here.