Abstract: Systems and methods for training object detection models using adversarial examples are provided. A method includes obtaining a training scene and identifying a target object within the training scene. The method includes obtaining an adversarial object and generating a modified training scene based on the adversarial object, the target object, and the training scene. The modified training scene includes the training scene modified to include the adversarial object placed on the target object. The modified training scene is input to a machine-learned model configured to detect the training object. A detection score is determined based on whether the training object is detected, and the machine-learned model and the parameters of the adversarial object are trained based on the detection output. The machine-learned model is trained to maximize the detection output. The parameters of the adversarial object are trained to minimize the detection output.

Abstract: Example aspects of the present disclosure describe a scene generator for simulating scenes in an environment. For example, snapshots of simulated traffic scenes can be generated by sampling a joint probability distribution trained on real-world traffic scenes. In some implementations, samples of the joint probability distribution can be obtained by sampling a plurality of factorized probability distributions for a plurality of objects for sequential insertion into the scene.

Abstract: Systems and methods for facilitating communication with autonomous vehicles are provided. In one example embodiment, a computing system can obtain a first type of sensor data (e.g., camera image data) associated with a surrounding environment of an autonomous vehicle and/or a second type of sensor data (e.g., LIDAR data) associated with the surrounding environment of the autonomous vehicle. The computing system can generate overhead image data indicative of at least a portion of the surrounding environment of the autonomous vehicle based at least in part on the first and/or second types of sensor data. The computing system can determine one or more lane boundaries within the surrounding environment of the autonomous vehicle based at least in part on the overhead image data indicative of at least the portion of the surrounding environment of the autonomous vehicle and a machine-learned lane boundary detection model.

Abstract: A method includes receiving image data associated with an image of a roadway including a crosswalk, generating a plurality of different characteristics of the image based on the image data, determining a position of the crosswalk on the roadway based on the plurality of different characteristics, the position including a first boundary and a second boundary of the crosswalk in the roadway, and providing map data associated with a map of the roadway, the map data including the position of the crosswalk on the roadway in the map. The plurality of different characteristics include a classification of one or more elements of the image, a segmentation of the one or more elements of the image, and one or more angles of the one or more elements of the image with respect to a line in the roadway.

Abstract: The present disclosure is directed encoding LIDAR point cloud data. In particular, a computing system can receive point cloud data for a three-dimensional space. The computing system can generate a tree-based data structure from the point cloud data, the tree-based data structure comprising a plurality of nodes. The computing system can generate a serial representation of the tree-based data structure. The computing system can, for each respective node represented by a symbol in the serial representation: determine contextual information for the respective node, generate, using the contextual information as input to a machine-learned model, a statistical distribution associated with the respective node, and generate a compressed representation of the symbol associated with the respective node by encoding the symbol using the statistical distribution for the respective node.

Abstract: Systems and methods are provided for machine-learned models including convolutional neural networks that generate predictions using continuous convolution techniques. For example, the systems and methods of the present disclosure can be included in or otherwise leveraged by an autonomous vehicle. In one example, a computing system can perform, with a machine-learned convolutional neural network, one or more convolutions over input data using a continuous filter relative to a support domain associated with the input data, and receive a prediction from the machine-learned convolutional neural network. A machine-learned convolutional neural network in some examples includes at least one continuous convolution layer configured to perform convolutions over input data with a parametric continuous kernel.

Abstract: Systems and methods for performing semantic segmentation of three-dimensional data are provided. In one example embodiment, a computing system can be configured to obtain sensor data including three-dimensional data associated with an environment. The three-dimensional data can include a plurality of points and can be associated with one or more times. The computing system can be configured to determine data indicative of a two-dimensional voxel representation associated with the environment based at least in part on the three-dimensional data. The computing system can be configured to determine a classification for each point of the plurality of points within the three-dimensional data based at least in part on the two-dimensional voxel representation associated with the environment and a machine-learned semantic segmentation model. The computing system can be configured to initiate one or more actions based at least in part on the per-point classifications.

Abstract: Generally, the disclosed systems and methods implement improved detection of objects in three-dimensional (3D) space. More particularly, an improved 3D object detection system can exploit continuous fusion of multiple sensors and/or integrated geographic prior map data to enhance effectiveness and robustness of object detection in applications such as autonomous driving. In some implementations, geographic prior data (e.g., geometric ground and/or semantic road features) can be exploited to enhance three-dimensional object detection for autonomous vehicle applications. In some implementations, object detection systems and methods can be improved based on dynamic utilization of multiple sensor modalities. More particularly, an improved 3D object detection system can exploit both LIDAR systems and cameras to perform very accurate localization of objects within three-dimensional space relative to an autonomous vehicle.

Abstract: Systems and methods for generating motion forecast data for actors with respect to an autonomous vehicle and training a machine learned model for the same are disclosed. The computing system can include an object detection model and a graph neural network including a plurality of nodes and a plurality of edges. The computing system can be configured to input sensor data into the object detection model; receive object detection data describing the location of the plurality of the actors relative to the autonomous vehicle as an output of the object detection model; input the object detection data into the graph neural network; iteratively update a plurality of node states respectively associated with the plurality of nodes; and receive, as an output of the graph neural network, the motion forecast data with respect to the plurality of actors.

Abstract: Systems and methods for answering region specific questions are provided. A method includes obtaining a regional scene question including an attribute query and a spatial region of interest for a training scene depicting a surrounding environment of a vehicle. The method includes obtaining a universal embedding for the training scene and an attribute embedding for the attribute query of the scene question. The universal embedding can identify sensory data corresponding to the training scene that can be used to answer questions concerning a number of different attributes in the training scene. The attribute embedding can identify aspects of an attribute that can be used to answer questions specific to the attribute. The method includes determining an answer embedding based on the universal embedding and the attribute embedding and determining a regional scene answer to the regional scene question based on the spatial region of interest and the answer embedding.

Abstract: Systems and methods of the present disclosure provide an improved approach for open-set instance segmentation by identifying both known and unknown instances in an environment. For example, a method can include receiving sensor point cloud input data including a plurality of three-dimensional points. The method can include determining a feature embedding and at least one of an instance embedding, class embedding, and/or background embedding for each of the plurality of three-dimensional points. The method can include determining a first subset of points associated with one or more known instances within the environment based on the class embedding and the background embedding associated with each point in the plurality of points. The method can include determining a second subset of points associated with one or more unknown instances within the environment based on the first subset of points. The method can include segmenting the input data into known and unknown instances.

Abstract: Example aspects of the present disclosure describe a scene generator for simulating scenes in an environment. For example, snapshots of simulated traffic scenes can be generated by sampling a joint probability distribution trained on real-world traffic scenes. In some implementations, samples of the joint probability distribution can be obtained by sampling a plurality of factorized probability distributions for a plurality of objects for sequential insertion into the scene.

Abstract: The present disclosure provides systems and methods that combine physics-based systems with machine learning to generate synthetic LiDAR data that accurately mimics a real-world LiDAR sensor system. In particular, aspects of the present disclosure combine physics-based rendering with machine-learned models such as deep neural networks to simulate both the geometry and intensity of the LiDAR sensor. As one example, a physics-based ray casting approach can be used on a three-dimensional map of an environment to generate an initial three-dimensional point cloud that mimics LiDAR data. According to an aspect of the present disclosure, a machine-learned geometry model can predict one or more adjusted depths for one or more of the points in the initial three-dimensional point cloud, thereby generating an adjusted three-dimensional point cloud which more realistically simulates real-world LiDAR data.

Abstract: Provided are systems and methods that perform multi-task and/or multi-sensor fusion for three-dimensional object detection in furtherance of, for example, autonomous vehicle perception and control. In particular, according to one aspect of the present disclosure, example systems and methods described herein exploit simultaneous training of a machine-learned model ensemble relative to multiple related tasks to learn to perform more accurate multi-sensor 3D object detection. For example, the present disclosure provides an end-to-end learnable architecture with multiple machine-learned models that interoperate to reason about 2D and/or 3D object detection as well as one or more auxiliary tasks. According to another aspect of the present disclosure, example systems and methods described herein can perform multi-sensor fusion (e.g.

Abstract: Systems, methods, tangible non-transitory computer-readable media, and devices associated with motion flow estimation are provided. For example, scene data including representations of an environment over a first set of time intervals can be accessed. Extracted visual cues can be generated based on the representations and machine-learned feature extraction models. At least one of the machine-learned feature extraction models can be configured to generate a portion of the extracted visual cues based on a first set of the representations of the environment from a first perspective and a second set of the representations of the environment from a second perspective. The extracted visual cues can be encoded using energy functions. Three-dimensional motion estimates of object instances at time intervals subsequent to the first set of time intervals can be determined based on the energy functions and machine-learned inference models.

Abstract: Systems and methods for generating object segmentations across videos are provided. An example system can enable an annotator to identify objects within a first image frame of a video sequence by clicking anywhere within the object. The system processes the first image frame and a second, subsequent, image frame to assign each pixel of the second image frame to one of the objects identified in the first image frame or the background. The system refines the resulting object masks for the second image frame using a recurrent attention module based on contextual features extracted from the second image frame. The system receives additional user input for the second image frame and uses the input, in combination with the object masks for the second image frame, to determine object masks for a third, subsequent, image frame in the video sequence. The process is repeated for each image in the video sequence.

Abstract: Systems and methods are provided for machine-learned models including convolutional neural networks that generate predictions using continuous convolution techniques. For example, the systems and methods of the present disclosure can be included in or otherwise leveraged by an autonomous vehicle. In one example, a computing system can perform, with a machine-learned convolutional neural network, one or more convolutions over input data using a continuous filter relative to a support domain associated with the input data, and receive a prediction from the machine-learned convolutional neural network. A machine-learned convolutional neural network in some examples includes at least one continuous convolution layer configured to perform convolutions over input data with a parametric continuous kernel.

Abstract: Systems, methods, tangible non-transitory computer-readable media, and devices associated with object perception and prediction of object motion are provided. For example, a plurality of temporal instance representations can be generated. Each temporal instance representation can be associated with differences in the appearance and motion of objects over past time intervals. Past paths and candidate paths of a set of objects can be determined based on the temporal instance representations and current detections of objects. Predicted paths of the set of objects using a machine-learned model trained that uses the past paths and candidate paths to determine the predicted paths. Past path data that includes information associated with the predicted paths can be generated for each object of the set of objects respectively.

Abstract: The present disclosure provides systems and methods for generating photorealistic image simulation data with geometry-aware composition for testing autonomous vehicles. In particular, aspects of the present disclosure can involve the intake of data on an environment and output of augmented data on the environment with the photorealistic addition of an object. As one example, data on the driving experiences of a self-driving vehicle can be augmented to add another vehicle into the collected environment data. The augmented data may then be used to test safety features of software for a self-driving vehicle.

Abstract: The present disclosure provides systems and methods that combine physics-based systems with machine learning to generate synthetic LiDAR data that accurately mimics a real-world LiDAR sensor system. In particular, aspects of the present disclosure combine physics-based rendering with machine-learned models such as deep neural networks to simulate both the geometry and intensity of the LiDAR sensor. As one example, a physics-based ray casting approach can be used on a three-dimensional map of an environment to generate an initial three-dimensional point cloud that mimics LiDAR data. According to an aspect of the present disclosure, a machine-learned model can predict one or more dropout probabilities for one or more of the points in the initial three-dimensional point cloud, thereby generating an adjusted three-dimensional point cloud which more realistically simulates real-world LiDAR data.