Abstract: Machine learning systems and methods that learn glare, and thus determine gaze direction in a manner more resilient to the effects of glare on input images. The machine learning systems have an isolated representation of glare, e.g., information on the locations of glare points in an image, as an explicit input, in addition to the image itself. In this manner, the machine learning systems explicitly consider glare while making a determination of gaze direction, thus producing more accurate results for images containing glare.

Abstract: An online system builds a high definition (HD) map for a geographical region based on sensor data captured by a plurality of autonomous vehicles driving through a geographical region. The autonomous vehicles detect map discrepancies based on differences in the surroundings observed using sensor data compared to the high definition map and send messages describing these map discrepancies to the online system. The online system updates existing occupancy maps to improve the accuracy of the occupancy maps (OMaps), and to thereby improve passenger and pedestrian safety. While vehicles are in motion, they can continuously collect data about their surroundings. When new data is available from the various vehicles within a fleet, this can be updated in a local representation of the occupancy map and can be passed to the online HD map system (e.g., in the cloud) for updating the master occupancy map shared by all of the vehicles.

Abstract: In training a deep neural network using reduced precision, gradient computation operates on larger values without affecting the rest of the training procedure. One technique trains the deep neural network to develop loss, scales the loss, computes gradients at a reduced precision, and reduces the magnitude of the computed gradients to compensate for scaling of the loss. In one example non-limiting arrangement, the training forward pass scales a loss value by some factor S and the weight update reduces the weight gradient contribution by 1/S. Several techniques can be used for selecting scaling factor S and adjusting the weight update.

Abstract: Vision transformers are deep learning models that employ a self-attention mechanism to obtain feature representations for an input image. To date, the configuration of vision transformers has limited the self-attention computation to a local window of the input image, such that short-range dependencies are modeled in the output. The present disclosure provides a vision transformer that captures global context, and that is therefore able to model long-range dependencies in its output.

Abstract: In various examples, a VPU and associated components may be optimized to improve VPU performance and throughput. For example, the VPU may include a min/max collector, automatic store predication functionality, a SIMD data path organization that allows for inter-lane sharing, a transposed load/store with stride parameter functionality, a load with permute and zero insertion functionality, hardware, logic, and memory layout functionality to allow for two point and two by two point lookups, and per memory bank load caching capabilities. In addition, decoupled accelerators may be used to offload VPU processing tasks to increase throughput and performance, and a hardware sequencer may be included in a DMA system to reduce programming complexity of the VPU and the DMA system. The DMA and VPU may execute a VPU configuration mode that allows the VPU and DMA to operate without a processing controller for performing dynamic region based data movement operations.

Abstract: While a compiler compiles source code to create an executable binary, code is added into the compiled source code that, when executed, identifies and stores in a metadata table base and bounds information associated with memory allocations. Additionally, additional code is added into the compiled source code that performs memory safety checks during execution. This updated compiled source code automatically determines a safety of memory access requests during execution by performing an out-of-bounds (OOB) check using the base and bounds information retrieved and stored in the metadata table. This enables the identification and avoidance of unsafe memory operations during the implementation of the executable by a GPU.

Abstract: Systems and methods relate to the determination of accurate motion vectors, for rendering situations such as a noisy Monte Carlo integration where image object surfaces are at least partially translucent. To optimize the search for “real world” positions, this invention defines the background as first path vertices visible through multiple layers of refractive interfaces. To find matching world positions, the background is treated as a single layer morphing in a chaotic way, permitting the optimized algorithm to be executed only once. Further improving performance over the prior linear gradient descent, the present techniques can apply a cross function and numerical optimization, such as Newton's quadratic target or other convergence function, to locate pixels via a vector angle minimization. Determined motion vectors can then serve as input for services including image denoising.

Abstract: Motivated by the ability of humans to quickly and accurately detect visual artifacts in images, a neural network model is trained to identify and locate visual artifacts in a sequence of rendered images without comparing the sequence of rendered images against a ground truth reference. Examples of visual artifacts include aliasing, blurriness, mosaicking, and overexposure. The neural network model provides a useful fully-automated tool for evaluating the quality of images produced by rendering systems. The neural network model may be trained to evaluate the quality of images for video processing, encoding, and/or compression techniques. In an embodiment, the sequence includes at least four images corresponding to a video or animation.

Abstract: In various examples, performance capabilities of a consumer—such as inference rates of a neural network executing on underlying hardware—may be tested or demonstrated for producers that have lower production rates than consumption rates of the consumer. For example, augmented data instances may be leveraged to augment production data from the producer in order to increase a frequency of transmission of data instances to a consumer. As a result, a data set including additional or alternative instances of data may be generated to emulate real-world data for demonstrating potential performance capabilities of a consuming algorithm beyond a currently required performance capability corresponding to a producer.

Abstract: Apparatuses, systems, and techniques to select fifth-generation (5G) new radio data. In at least one embodiment, a processor includes one or more circuits to select 5G new radio signal information in parallel.

Abstract: Camera based localization performed to determine a current pose of an autonomous vehicle without the aid of depth sensors such as LiDAR. The vehicle comprises an imaging system configured to capture image frames depicting portions of the surrounding area. Based on an initial pose of the vehicle, edgels corresponding to three-dimensional locations are loaded and mapped to corresponding edge pixels of the captured image frame. A pose of the vehicle is optimized based upon the determined correspondences by identifying a transformation that minimizes a distance between the edgels and their corresponding edge pixels. The determined transformation can be applied to the initial pose to determine an updated pose of the vehicle.

Abstract: Apparatuses, systems, and techniques to optimize memory usage when performing matrix operations. In at least one embodiment, a matrix is optimized to limit memory and storage requirements while minimizing loss of precision for a sum of the members of the matrix.

Abstract: In at least one embodiment, under the control of a robotic control system, a gripper on a robot is positioned to grasp a 3-dimensional object. In at least one embodiment, the relative position of the object and the gripper is determined, at least in part, by using a camera mounted on the gripper.

Abstract: Operations of the present disclosure include obtaining a first measure of velocity of a vehicle based on a plurality of locations determined for the vehicle. The operations also include obtaining, based on IMU measurements of an inertial measurement unit (IMU) of the vehicle, a second measure of velocity of the vehicle. In addition, the operations include performing calibration of the IMU based on the first measure of velocity and the second measure of velocity.

Abstract: An electronic device includes an integrated circuit and a heat exchanger. The heat exchanger includes a heat pipe and a first plurality of cooling fins and a second plurality of cooling fins. The heat pipe is thermally coupled to the integrated circuit and has an evaporator portion and a condenser portion, where the condenser portion extends away from the evaporator portion. The first plurality of cooling fins are attached to the condenser portion and proximate to the evaporation portion and form a plenum having a first associated pressure drop when a cooling fluid flows across the first plurality of cooling fins at a first velocity. The second plurality of cooling fins are attached to the condenser portion and distal from the evaporation portion and form a flow path having a second associated pressure drop when the cooling fluid flows across the second plurality of cooling fins at the first velocity.

Abstract: A manifold voxel mesh or surface mesh is manufacturable by carving a single block of material and a non-manifold mesh is not manufacturable. Conventional techniques for constructing or extracting a surface mesh from an input point cloud often produce a non-manifold voxel mesh. Similarly, extracting a surface mesh from a voxel mesh that includes non-manifold geometry produces a surface mesh that includes non-manifold geometry. To ensure that the surface mesh includes only manifold geometry, locations of the non-manifold geometry in the voxel mesh are detected and converted into manifold geometry. The result is a manifold voxel mesh from which a manifold surface mesh of the object may be extracted.

Abstract: State information can be determined for a subject that is robust to different inputs or conditions. For drowsiness, facial landmarks can be determined from captured image data and used to determine a set of blink parameters. These parameters can be used, such as with a temporal network, to estimate a state (e.g., drowsiness) of the subject. To improve robustness, an eye state determination network can determine eye state from the image data, without reliance on intermediate landmarks, that can be used, such as with another temporal network, to estimate the state of the subject. A weighted combination of these values can be used to determine an overall state of the subject. To improve accuracy, individual behavior patterns and context information can be utilized to account for variations in the data due to subject variation or current context rather than changes in state.

Abstract: In various examples, a single camera is used to capture two images of a scene from different locations. A trained neural network, taking the two images as inputs, outputs a scene structure map that indicates a ratio of height and depth values for pixel locations associated with the images. This ratio may indicate the presence of an object above a surface (e.g., road surface) within the scene. Object detection then can be performed on non-zero values or regions within the scene structure map.

Abstract: One embodiment of a computer-implemented method for processing data within a fixed-function pipeline included in an execution unit includes receiving a first input from a first processing unit, wherein the first input corresponds to a first fixed-function; executing the first fixed-function on the first input to generate a first output, wherein the first fixed-function is executed on the first input prior to executing the first fixed-function on one or more inputs received from a plurality of processing cores that are processing a plurality of rays, and wherein each ray represents a path from a light source to at least one point within a three-dimensional (3D) environment; and transmitting the first output to the first processing unit for further processing.

Abstract: Systems and methods are disclosed related to low-power dynamic offset calibration of an error amplifier. An analog linear voltage regulator circuit tracks changes between a reference voltage and a regulated voltage to keep the regulated voltage as close as possible to the reference voltage. The analog linear voltage regulator includes an error amplifier that measures the error between the reference and regulated voltages and feedback circuitry. The error amplifier and feedback circuitry should be calibrated to correct for any offset within the circuits. The described offset calibration technique not only compensates for the offset in the error amplifier but also cancels any mismatch in the feedback network. During operation, conditions such as temperature and supply voltage may vary causing the offset to change. The technique is low power and dynamically cancels the offset even when the linear regulator is operating to supply the desired voltage.