Magic Leap, Inc.

Google Inc.

Gm Cruise Holdings LLC

Zhao Chen

Vijay Badrinarayanan

Andrew Rabinovich

Ayan Tuhinendu Sinha

Anelia Angelova

Yuning Chai

Elad Joseph

Soeren Pirk

Ariel Gordon

Yao Lu

Vincent Michael Casser

Ameya Pramod Phalak

Anthony Brohan

Jan Dlabal

Hongge Chen

The present disclosure provides solutions for generating synthetic three-dimensional (3D) objects for testing autonomous vehicle (AV) object detectors, such as AV Light Detection and Ranging (LiDAR) detectors. A process of the disclosed technology can include steps for receiving a first set of object shape data corresponding to a baseline object, permuting the first set of object shape data to generate a second set of object shape data, and providing the second set of object shape data to a differentiable pipeline to generate an output scene, wherein the output scene contains an optimized object that is based on the second set of object shape data. The process can further include steps for determining, using a detector, if the optimized object can be identified in the output scene. Systems and machine-readable media are also provided.

Pipeline for generating synthetic point cloud data

Systems and methods for gradient adversarial training of a neural network are disclosed. In one aspect of gradient adversarial training, an auxiliary neural network can be trained to classify a gradient tensor that is evaluated during backpropagation in a main neural network that provides a desired task output. The main neural network can serve as an adversary to the auxiliary network in addition to a standard task-based training procedure. The auxiliary neural network can pass an adversarial gradient signal back to the main neural network, which can use this signal to regularize the weight tensors in the main neural network. Gradient adversarial training of the neural network can provide improved gradient tensors in the main network. Gradient adversarial techniques can be used to train multitask networks, knowledge distillation networks, and adversarial defense networks.

Gradient adversarial training of neural networks

Systems and methods are disclosed for training and using neural networks for computing depth maps. One method for training the neural network includes providing an image input to the neural network. The image input may include a camera image of a training scene. The method may also include providing a depth input to the neural network. The depth input may be based on a high-density depth map of the training scene and a sampling mask. The method may further include generating, using the neural network, a computed depth map of the training scene based on the image input and the depth input. The method may further include modifying the neural network based on an error between the computed depth map and the high-density depth map.

Image-enhanced depth sensing using machine learning

Techniques for estimating a perimeter of a room environment at least partially enclosed by a set of adjoining walls using posed images are disclosed. A set of images and a set of poses are obtained. A depth map is generated based on the set of images and the set of poses. A set of wall segmentation maps are generated based on the set of images, each of the set of wall segmentation maps indicating a target region of a corresponding image that contains the set of adjoining walls. A point cloud is generated based on the depth map and the set of wall segmentation maps, the point cloud including a plurality of points that are sampled along portions of the depth map that align with the target region. The perimeter of the environment along the set of adjoining walls is estimated based on the point cloud.

Perimeter estimation from posed monocular video

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a neural network using consistency measures. One of the methods includes processing a particular training example from a mediator training data set using a first neural network to generate a first output for a first machine learning task; processing the particular training example in the mediator training data set using each of one or more second neural networks, wherein each second neural network is configured to generate a second output for a respective second machine learning task; determining, for each second machine learning task, a consistency target output for the first machine learning task; determining, for each second machine learning task, an error between the first output and the consistency target output corresponding to the second machine learning task; and generating a parameter update for the first neural network from the determined errors.

Training neural networks using consistency measures

Systems and methods for training a multitask network is disclosed. In one aspect, training the multitask network includes determining a gradient norm of a single-task loss adjusted by a task weight for each task, with respect to network weights of the multitask network, and a relative training rate for the task based on the single-task loss for the task. Subsequently, a gradient loss function, comprising a difference between (1) the determined gradient norm for each task and (2) a corresponding target gradient norm, can be determined. An updated task weight for the task can be determined and used in the next iteration of training the multitask network, using a gradient of the gradient loss function with respect to the task weight for the task.

Gradient normalization systems and methods for adaptive loss balancing in deep multitask networks

Systems and methods are disclosed for computing depth maps. One method includes capturing, using a camera, a camera image of a runtime scene. The method may also include analyzing the camera image of the runtime scene to determine a plurality of target sampling points at which to capture depth of the runtime scene. The method may further include adjusting a setting associated with a low-density depth sensor based on the plurality of target sampling points. The method may further include capturing, using the low-density depth sensor, a low-density depth map of the runtime scene at the plurality of target sampling points. The method may further include generating a computed depth map of the runtime scene based on the camera image of the runtime scene and the low-density depth map of the runtime scene.

Growing community of inventors

Mountain View, CA, United States of America

Zhao Chen