Toyota Research Institute, Inc.

Toyota Jidosha Kabushiki Kaisha

Leland Stanford Junior University

Toyota Technological Institute at Chicago

Massachusetts Institute of Technology

The Board of Trustees of the Leland Standford Junior Univeristy

Rares A Ambrus

Adrien David Gaidon

Vitor Guizilini

Sudeep Pillai

Jie Li

Igor Vasiljevic

Sergey Zakharov

Jiexiong Tang

Dennis Park

Gregory Shakhnarovich

Hanme Kim

Victor Vaquero Gomez

Kuan-Hui Lee

Arjun Bhargava

Leonidas John Guibas

Wolfram Burgard

Or Litany

Jiading Fang

Matthew R Walter

Christin Jeannette Bohg

Haofeng Chen

Guy Rosman

Alexander Alspach

Naveen Suresh Kuppuswamy

Andrew M Beaulieu

Fredo Durand

Jia-En M Pan

Pavel V Tokmakov

Xipeng Wang

Vitor Campagnolo Guizilini

Rui Hou

Vincent Sitzmann

Dian Chen

Wadim Kehl

Joshua Tenenbaum

Jiajun Wu

Jeff Walls, Sr

Taraneh Sadjadpour

Andreas Buehler

Greg Shakhnarovich

Yue Wang

Hsu-kuang Chiu

Xiangru Huang

Justin Solomon

Systems and methods for performing three-dimensional multi-object tracking are disclosed herein. In one example, a method includes the steps of determining a residual based on augmented current frame detection bounding boxes, augmented previous frame detection bounding boxes, augmented current frame shape descriptors, and augmented previous frame shape descriptors and predicting an affinity matrix using the residual. The residual indicates a spatiotemporal and shape similarity between current detections in a current frame point cloud data and previous detections in a previous frame point cloud data. The affinity matrix indicates associations between the previous detections and the current detections, as well as the augmented anchors.

System and method for three-dimensional multi-object tracking

Systems and methods for enhanced computer vision capabilities, particularly including depth synthesis, which may be applicable to autonomous vehicle operation are described. A vehicle may be equipped with a geometric scene representation (GSR) architecture for synthesizing depth views at arbitrary viewpoints. The GSR architecture synthesizes depth views enable advanced functions, including depth interpolation and depth extrapolation. The GSR architecture implements functions (i.e., depth interpolation, depth extrapolation) that are useful for various computer vision applications for autonomous vehicles, such as predicting depth maps from unseen locations. For example, a vehicle includes a processor device synthesizing depth views at multiple viewpoints, where the multiple viewpoints are from image data of a surrounding environment for the vehicle. Further, the vehicle can have a controller device that receives depth views from the processor device and performs autonomous operations in response to analysis of the depth views.

Systems and methods for depth synthesis with transformer architectures

A method of representation learning for object detection from unlabeled point cloud sequences is described. The method includes detecting moving object traces from temporally-ordered, unlabeled point cloud sequences. The method also includes extracting a set of moving objects based on the moving object traces detected from the sequence of temporally-ordered, unlabeled point cloud sequences. The method further includes classifying the set of moving objects extracted from on the moving object traces detected from the sequence of temporally-ordered, unlabeled point cloud sequences. The method also includes estimating 3D bounding boxes for the set of moving objects based on the classifying of the set of moving objects.

Representation learning for object detection from unlabeled point cloud sequences

A method for improving 3D object detection via object-level augmentations is described. The method includes recognizing, using an image recognition model of a differentiable data generation pipeline, an object in an image of a scene. The method also includes generating, using a 3D reconstruction model, a 3D reconstruction of the scene from the image including the recognized object. The method further includes manipulating, using an object level augmentation model, a random property of the object by a random magnitude at an object level to determine a set of properties and a set of magnitudes of an object manipulation that maximizes a loss function of the image recognition model. The method also includes training a downstream task network based on a set of training data generated based on the set of properties and the set of magnitudes of the object manipulation, such that the loss function is minimized.

Adversarial object-aware neural scene rendering for 3D object detection

System, methods, and other embodiments described herein relate to an improved approach to training a depth model for monocular depth estimation by warping depth features prior to decoding. In one embodiment, a method includes encoding, using an encoder of a depth model, a source image into depth features of a scene depicted by the source image. The method includes warping the depth features into warped features of a target frame of a target image associated with the source image. The method includes decoding, using a decoder of the depth model, the warped features into a depth map. The method includes training the depth model according to a loss derived from the depth map.

Warping depth features for depth estimation

System, methods, and other embodiments described herein relate to an improved approach to training a depth model to derive depth estimates from monocular images using histograms to assess photometric losses. In one embodiment, a method includes determining loss values according to a photometric loss function. The loss values are associated with a depth map derived from an input image that is a monocular image. The method includes generating histograms for the loss values corresponding to different regions of a target image. The method includes, responsive to identifying erroneous values of the loss values, masking the erroneous values to avoid considering the erroneous values during training of the depth model.

Using histograms for self-supervised depth estimation

System, methods, and other embodiments described herein relate to an improved approach to training a depth model to derive depth estimates from monocular images using cost volumes. In one embodiment, a method includes predicting, using a depth model, depth values from at least one input image that is a monocular image. The method includes generating a cost volume by sampling the depth values corresponding to bins of the cost volume. The method includes determining loss values for the bins of the cost volume. The method includes training the depth model according to the loss values of the cost volume.

Photometric cost volumes for self-supervised depth estimation

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

A method for 3D object detection is described. The method includes concurrently training a monocular depth network and a 3D object detection network. The method also includes predicting, using a trained monocular depth network, a monocular depth map of a monocular image of a video stream. The method further includes inferring a 3D point cloud of a 3D object within the monocular image according to the predicted monocular depth map. The method also includes predicting 3D bounding boxes from a selection of 3D points from the 3D point cloud of the 3D object based on a selection regression loss.

Monocular object detection via end-to-end differentiable pipeline

System, methods, and other embodiments described herein relate to training a depth model for monocular depth estimation using photometric loss masks derived from motion estimates of dynamic objects. In one embodiment, a method includes generating depth maps from images of an environment. The method includes determining motion of points within the depth maps. The method includes associating the points between the depth maps to identify an object according to a correlation of the motion for a first cluster of the points with a second cluster of the points. The method includes providing the depth maps and the object as an electronic output.

Self-supervised monocular depth estimation via rigid-motion embeddings

A method for controlling an agent to navigate through an environment includes generating a depth map associated with a monocular image of the environment. The method also includes generating a group of surface normal. Each surface normal of the group of surface normals is associated with a respective polygon of a group of polygons associated with the depth map. The method further includes identifying one or more ground planes in the depth map based on the group of surface normal. The method further includes controlling the agent to navigate through the environment based on identifying the one or more ground planes.

System and method for self-supervised monocular ground-plane extraction

Systems, methods, and other embodiments described herein relate to estimating scaled depth maps by sampling variational representations of an image using a learning model. In one embodiment, a method includes encoding data embeddings by a learning model to form conditioned latent representations using attention networks, the data embeddings including features about an image from a camera and calibration information about the camera. The method also includes computing a probability distribution of the conditioned latent representations by factoring scale priors. The method also includes sampling the probability distribution to generate variations for the data embeddings. The method also includes estimating scaled depth maps of a scene from the variations at different coordinates using the attention networks.

Systems and methods for estimating scaled maps by sampling representations from a learning model

A method for depth estimation performed by a depth estimation system associated with an agent includes determining a first depth of a first image and a second depth of a second image, the first image and the second image being captured by a sensor associated with the agent. The method also includes generating a first 3D image of the first image based on the first depth, a first pose associated with the sensor, and the second image. The method further includes generating a warped depth image based on transforming the first depth in accordance with the first pose. The method also includes updating the first pose based on a second pose associated with the warped depth image and the second depth, and updating the first 3D image based on the updated first pose. The method further includes controlling an action of the agent based on the updated first 3D image.

Depth estimation based on ego-motion estimation and residual flow estimation

A system for producing a depth map can include a processor and a memory. The memory can store a neural network. The neural network can include an encoding portion module, a multi-frame feature matching portion module, and a decoding portion module. The encoding portion module can include instructions that, when executed by the processor, cause the processor to encode an image to produce single-frame features. The multi-frame feature matching portion module can include instructions that, when executed by the processor, cause the processor to process the single-frame features to produce information. The decoding portion module can include instructions that, when executed by the processor, cause the processor to decode the information to produce the depth map. A first training dataset, used to train the multi-frame feature matching portion module, can be different from a second training dataset used to train the encoding portion module and the decoding portion module.

Producing a depth map from a monocular two-dimensional image

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

Growing community of inventors

San Francisco, CA, United States of America

Rares A Ambrus