Abstract: A neural network system, comprising: instructions for implementing at least a SWBN layer in a neural network, and wherein the instructions perform operations comprising: during training of the neural network system on a plurality of batches of training data and for each of the plurality of batches: obtaining a respective first layer output for each of the plurality of training data; determining a plurality of normalization statistics for the batch from the first layer outputs; generating a respective normalized output for each training data in the batch; updating the whitening matrix by a covariance matrix; performing stochastic whitening on the normalized components of each first layer output; transforming the whitened data for each training data; generating a respective SWBN layer output for each of the training data from the transformed whitened data for each training data in the batch; and providing the SWBN layer output.

Abstract: Methods and systems for deep neural networks using dynamically selected feature-relevant points from a point cloud are described. A plurality of multidimensional feature vectors arranged in a point-feature matrix are received. Each row of the point-feature matrix corresponds to a respective one of the multidimensional feature vectors, and each column of the point-feature matrix corresponds to a respective feature. Each multidimensional feature vector represents a respective unordered point from a point cloud and each multidimensional feature vector includes a respective plurality of feature-correlated values, each feature-correlated value represents a correlation extent of the respective feature. A reduced-max matrix having a selected plurality of feature-relevant vectors is generated.

Abstract: A neural network system, comprising: instructions for implementing at least a SWBN layer in a neural network, and wherein the instructions perform operations comprising: during training of the neural network system on a plurality of batches of training data and for each of the plurality of batches: obtaining a respective first layer output for each of the plurality of training data; determining a plurality of normalization statistics for the batch from the first layer outputs; generating a respective normalized output for each training data in the batch; updating the whitening matrix by a covariance matrix; performing stochastic whitening on the normalized components of each first layer output; transforming the whitened data for each training data; generating a respective SWBN layer output for each of the training data from the transformed whitened data for each training data in the batch; and providing the SWBN layer output.

Abstract: Methods and apparatuses for generating a frame of semantically labeled 2D data are described. A frame of sparse 3D data is generated from a frame of sparse 3D data. Semantic labels are assigned to the frame of dense 3D data, based on a set of 3D bounding boxes determined for the frame of sparse 3D data. Semantic labels are assigned to a corresponding frame of 2D data based on a mapping between the frame of sparse 3D data and the frame of 2D data. The mapping is used to map a 3D data point in the frame of dense 3D data to a mapped 2D data point in the frame of 2D data. The semantic label assigned to the 3D data point is assigned to the mapped 2D data point. The frame of semantically labeled 2D data, including the assigned semantic labels, is outputted.

Abstract: Upon receiving a set of two-dimensional data points representing an object in an environment, a bounding box estimator estimates a bounding box vector representative of a two-dimensional version of the object that is represented by the two-dimensional data points.

Abstract: Methods and systems for encoding 3D data for use with 2D convolutional neural networks (CNNs) are described. A set of 3D data is encoded into a set of one or more arrays. A 2D index of the arrays is calculated by projecting 3D coordinates of the 3D point onto a 2D image plane that is defined by a set of defined virtual camera parameters. The virtual camera parameters include a camera projection matrix defining the 2D image plane. Each 3D coordinate of the point is stored in the arrays at the calculated 2D index. The set of encoded arrays is provided for input to a 2D CNN, for training or inference.

Abstract: Methods and apparatuses for generating a frame of semantically labeled 2D data are described. A frame of sparse 3D data is generated from a frame of sparse 3D data. Semantic labels are assigned to the frame of dense 3D data, based on a set of 3D bounding boxes determined for the frame of sparse 3D data. Semantic labels are assigned to a corresponding frame of 2D data based on a mapping between the frame of sparse 3D data and the frame of 2D data. The mapping is used to map a 3D data point in the frame of dense 3D data to a mapped 2D data point in the frame of 2D data. The semantic label assigned to the 3D data point is assigned to the mapped 2D data point. The frame of semantically labeled 2D data, including the assigned semantic labels, is outputted.

Abstract: Methods and systems for deep neural networks using dynamically selected feature-relevant points from a point cloud are described. A plurality of multidimensional feature vectors arranged in a point-feature matrix are received. Each row of the point-feature matrix corresponds to a respective one of the multidimensional feature vectors, and each column of the point-feature matrix corresponds to a respective feature. Each multidimensional feature vector represents a respective unordered point from a point cloud and each multidimensional feature vector includes a respective plurality of feature-correlated values, each feature-correlated value represents a correlation extent of the respective feature. A reduced-max matrix having a selected plurality of feature-relevant vectors is generated.

Abstract: Methods and systems for encoding 3D data for use with 2D convolutional neural networks (CNNs) are described. A set of 3D data is encoded into a set of one or more arrays. A 2D index of the arrays is calculated by projecting 3D coordinates of the 3D point onto a 2D image plane that is defined by a set of defined virtual camera parameters. The virtual camera parameters include a camera projection matrix defining the 2D image plane. Each 3D coordinate of the point is stored in the arrays at the calculated 2D index. The set of encoded arrays is provided for input to a 2D CNN, for training or inference.

Abstract: Upon receiving a set of two-dimensional data points representing an object in an environment, a bounding box estimator estimates a bounding box vector representative of a two-dimensional version of the object that is represented by the two-dimensional data points.