Toyota Research Institute, Inc.

Sudeep Pillai

Adrien David Gaidon

Rares A Ambrus

Vitor Guizilini

Kun-Hsin Chen

Arjun Bhargava

Kuan-Hui Lee

Igor Vasiljevic

Jie Li

Hanme Kim

Shunsho Kaku

Hai Jin

Peiyan Gong

Jiexiong Tang

Jia-En M Pan

Haofeng Chen

Chao Fang

James Joseph Kuffner

Ryan W Wolcott

Nikos Arechiga Gonzalez

Daniele Molinari

Logan Michael Ellis

Wolfram Burgard

Xipeng Wang

Constantin Franziskus Dominik Hubmann

Kuan-hui Lee

Jeff Walls, Sr

T Wolfram Burgard

Daniel D Kang

Allan Raventos

Sarah Yoo

David L Garber

A method for semantic keypoint detection is described. The method includes linking, using a keypoint graph neural network (KGNN), semantic keypoints of an object within a first image of a video stream into a 2D graph structure corresponding to a category of the object. The method also includes embedding descriptors within the semantic keypoints of the 2D graph structure corresponding to the category of the object. The method further includes tracking the object within subsequent images of the video stream using the embedded descriptors within the semantic keypoints of the 2D graph structure corresponding to the category of the object.

End-to-end monocular 2D semantic keypoint detector and tracker learning

Systems, methods, computer-readable media, techniques, and methodologies are disclosed for performing end-to-end, learning-based keypoint detection and association. A scene graph of a signalized intersection is constructed from an input image of the intersection. The scene graph includes detected keypoints and linkages identified between the keypoints. The scene graph can be used along with a vehicle's localization information to identify which keypoint that represents a traffic signal is associated with the vehicle's current travel lane. An appropriate vehicle action may then be determined based on a transition state of the traffic signal keypoint and trajectory information for the vehicle. A control signal indicative of this vehicle action may then be output to cause an autonomous vehicle, for example, to implement the appropriate vehicle action.

End-to-end signalized intersection transition state estimator with scene graphs over semantic keypoints

Systems and methods for self-supervised learning for visual odometry using camera images, may include: estimating correspondences between keypoints of a target camera image and keypoints of a context camera image; based on the keypoint correspondences, lifting a set of 2D keypoints to 3D, using a neural camera model; and projecting the 3D keypoints into the context camera image using the neural camera model. Some embodiments may use the neural camera model to achieve the lifting and projecting of keypoints without a known or calibrated camera model.

Systems and methods for generic visual odometry using learned features via neural camera models

Systems and methods are provided for implementing soft model assertions (SMA) system and techniques designed to monitor and improve Machine Learning (ML) model quality by to detecting errors within the one or more ML models. SMA techniques and systems are distinctly designed to leverage: 1) a user's ability to specify features over data; and 2) large, existing datasets of organizations, in a manner that can improve the accuracy and quality of predicting potential errors in Machine Learning (ML) models. A SMA system can include a controller device receiving predictions generated based on the ML models and output from the SMA system. The controller performs autonomous operations of the system in response to determining that the one or more detected errors within the one or more ML models yield a high certainty of errors in the predictions. The SMA system also includes a domain specific language and a severity score module.

Systems and methods for soft model assertions

A method for 2D semantic keypoint detection and tracking is described. The method includes learning embedded descriptors of salient object keypoints detected in previous images according to a descriptor embedding space model. The method also includes predicting, using a shared image encoder backbone, salient object keypoints within a current image of a video stream. The method further includes inferring an object represented by the predicted, salient object keypoints within the current image of the video stream. The method also includes tracking the inferred object by matching embedded descriptors of the predicted, salient object keypoints representing the inferred object within the previous images of the video stream based on the descriptor embedding space model.

Monocular 2D semantic keypoint detection and tracking

A two dimensional image can be received. A depth map can be produced, via a first neural network, from the two dimensional image. A bird's eye view image can be produced, via a second neural network, from the depth map. The second neural network can implement a machine learning algorithm that preserves spatial gradient information associated with one or more objects included in the depth map and causes a position of a pixel in an object, included in the bird's eye view image, to be represented by a differentiable function. Three dimensional objects can be detected, via a third neural network, from the two dimensional image, the bird's eye view image, and the spatial gradient information. A combination of the first neural network, the second neural network, and the third neural network can be end-to-end trainable and can be included in a perception system.

Producing a bird's eye view image from a two dimensional image

Datasets for autonomous driving systems and multi-modal scenes may be automatically labeled using previously trained models as priors to mitigate the limitations of conventional manual data labeling. Properly versioned models, including model weights and knowledge of the dataset on which the model was previously trained, may be used to run an inference operation on unlabeled data, thus automatically labeling the dataset. The newly labeled dataset may then be used to train new models, including sparse data sets, in a semi-supervised or weakly-supervised fashion.

Systems and methods for dataset and model management for multi-modal auto-labeling and active learning

A method for estimating depth of a scene includes selecting an image of the scene from a sequence of images of the scene captured via an in-vehicle sensor of a first agent. The method also includes identifying previously captured images of the scene. The method further includes selecting a set of images from the previously captured images based on each image of the set of images satisfying depth criteria. The method still further includes estimating the depth of the scene based on the selected image and the selected set of images.

Multi-view depth estimation leveraging offline structure-from-motion

A method for learning depth-aware keypoints and associated descriptors from monocular video for monocular visual odometry is described. The method includes training a keypoint network and a depth network to learn depth-aware keypoints and the associated descriptors. The training is based on a target image and a context image from successive images of the monocular video. The method also includes lifting 2D keypoints from the target image to learn 3D keypoints based on a learned depth map from the depth network. The method further includes estimating a trajectory of an ego-vehicle based on the learned 3D keypoints.

Self-supervised 3D keypoint learning for monocular visual odometry

A vehicle light classification system captures a sequence of images of a scene that includes a front/rear view of a vehicle with front/rear-side lights, determines semantic keypoints, in the images and associated with the front/rear-side lights, based on inputting the images into a first neural network, obtains multiple difference images that are each a difference between successive images from among the sequence of images, the successive images being aligned based on their respective semantic keypoints, and determines a classification of the front/rear-side lights based at least in part on the difference images by inputting the difference images into a second neural network.

Systems and methods for vehicle light signal classification

A method for localization performed by an agent includes receiving a query image of a current environment of the agent captured by a sensor integrated with the agent. The method also includes receiving a target image comprising a first set of keypoints matching a second set of keypoints of the query image. The first set of keypoints may be generated based on a task specified for the agent. The method still further includes determining a current location based on the target image.

Incremental map building using learnable features and descriptors

A method for controlling an ego agent includes capturing a two-dimensional (2D) image of an environment adjacent to the ego agent. The method also includes generating a semantically segmented image of the environment based on the 2D image. The method further includes generating a depth map of the environment based on the semantically segmented image. The method additionally includes generating a three-dimensional (3D) estimate of the environment based on the depth map. The method also includes controlling an action of the ego agent based on the identified location.

System and method for monocular depth estimation from semantic information

Systems and methods are provided for developing/updating training datasets for traffic light detection/perception models. V2I-based information may indicate a particular traffic light state/state of transition. This information can be compared to a traffic light perception prediction. When the prediction is inconsistent with the V2I-based information, data regarding the condition(s)/traffic light(s)/etc. can be saved and uploaded to a training database to update/refine the training dataset(s) maintained therein. In this way, an existing traffic light perception model can be updated/improved and/or a better traffic light perception model can be developed.

Traffic light active learning pipeline

A method for 3D object tracking is described. The method includes inferring first 2D semantic keypoints of a 3D object within a sparsely annotated video stream. The method also includes matching the first 2D semantic keypoints of a current frame with second 2D semantic keypoints in a next frame of the sparsely annotated video stream using embedded descriptors within the current frame and the next frame. The method further includes warping the first 2D semantic keypoints to the second 2D semantic keypoints to form warped 2D semantic keypoints in the next frame. The method also includes labeling a 3D bounding box in the next frame according to the warped 2D semantic keypoints in the next frame.

Semi-supervised 3D object tracking in videos via 2D semantic keypoints

Systems and methods to improve machine learning by explicitly over-fitting environmental data obtained by an imaging system, such as a monocular camera are disclosed. The system includes training self-supervised depth and pose networks in monocular visual data collected from a certain area over multiple passes. Pose and depth networks may be trained by extracting data from multiple images of a single environment or trajectory, allowing the system to overfit the image data.

Growing community of inventors

Santa Clara, CA, United States of America

Sudeep Pillai