Abstract: Disclosed herein are systems and methods for performing post training quantization. A processor obtains fixed-point output values from a layer of an artificial neural network (ANN) wherein the layer includes fixed-point weights determined based on floating-point weights and a weight scaling factor determined based on an output scaling factor. Next, the processor converts the fixed-point output values to floating-point output values based on the output scaling factor. Then, the processor expands a range of floating-point values. Next, the processor calculates a new output scaling factor based on the expanded range of floating-point output values. Finally, the processor stores the new output scaling factor in an associated memory.

Abstract: This invention predicts that intra mode prediction is more effective for the macroblocks where motion estimation in inter mode prediction fails. This failure is indicated by a large value of the inter mode SAD. This invention performs intra mode prediction for only macro blocks have larger inter mode SADs. The definition of a large inter mode SAD differs for different content. This invention compares the inter mode SAD of a current macroblock with an adaptive threshold. This adaptive threshold depends on the average and variance of the SADs of the previous predicted frame. An adaptive threshold is calculated for each new predictive frame.

Abstract: In an example, a method includes executing, using one or more processors, a power-of-2 parametric activation (PACT2) function to quantize a set of data. The executing of the PACT2 function includes determining a distribution for the set of data; discarding a portion of the data corresponding to a tail of the distribution to form a remaining set of data; estimating a maximum value of the remaining set of data; determining a new maximum value of the remaining set of data using a moving average and at least one historical value of at least one prior remaining set of data; determining a clipping value by expanding the new maximum value to a nearest power of two value; and quantizing the set of data using the clipping value to form a quantized set of data.

Abstract: In described examples of a method for quantizing data for a convolutional neural network (CNN) is provided. A set of data is received and quantized the using a power-of-2 parametric activation (PACT2) function. The PACT2 function arranges the set of data as a histogram and discards a portion of the data corresponding to a tail of the histogram to form a remaining set of data. A clipping value is determined by expanding the remaining set of data to a nearest power of two value. The set of data is then quantized using the clipping value. With PACT2, a model can be quantized either using post training quantization or using quantization aware training. PACT2 helps a quantized model to achieve close accuracy compared to the corresponding floating-point model.

Abstract: Disclosed herein are systems and methods that provide an end-to-end approach for performing multi-dimensional object pose estimation in the context of machine learning models. In an implementation, processing circuitry of a suitable computer inputs image data to a machine learning model that predicts a parameterized rotation vector and a parameterized translation vector for an object in the image. Next, the processing circuitry converts the parameterized rotation vector and the parameterized translation vector into a non-parameterized rotation vector and a non-parameterized translation vector respectively. Finally, the processing circuitry updates the image data based on the non-parameterized rotation vector and the non-parameterized translation vector.

Abstract: A method for convolution in a convolutional neural network (CNN) is provided that includes accessing a coefficient value of a filter corresponding to an input feature map of a convolution layer of the CNN, and performing a block multiply accumulation operation on a block of data elements of the input feature map, the block of data elements corresponding to the coefficient value, wherein, for each data element of the block of data elements, a value of the data element is multiplied by the coefficient value and a result of the multiply is added to a corresponding data element in a corresponding output block of data elements comprised in an output feature map.

Abstract: Various examples disclosed herein relate to neural network quantization techniques, and more particularly, to selecting inference precisions for the layers of the neural network. In an example embodiment, a method is provided herein that includes determining an accuracy improvement of a layer of a neural network implemented using a first bit precision relative to using a second bit precision and determining a latency degradation of the layer of the neural network implemented using the first bit precision relative to using the second bit precision. The method further includes selecting, based on the accuracy improvement and the latency degradation, the first bit precision or the second bit precision for use in implementing the layer of the neural network.

Abstract: Disclosed herein are systems and methods for determining the scaling factors for a neural network that satisfy the activation functions employed by the nodes of the network. A processor identifies a saturation point of an activation function. Next, the processor determines a scaling factor for an output feature map based on the saturation point of the activation function. Then, the processor determines a scaling factor for an accumulator based on the scaling for the output feature map and further based on a shift value related to a quantization. Finally, the processor determines a scaling factor for a weight map based on the scaling factor for the accumulator.

Abstract: Techniques including receiving a distorted image from a camera disposed about a vehicle, detecting, in the distorted image, corner points associated with a target object, mapping the corner points to a distortion corrected domain based on one or more camera parameters, mapping the corner points and lines between the corner points back to a distorted domain based on the camera parameters, interpolating one or more intermediate points to generate lines between the corner points in the distortion corrected domain mapping the corner points and the lines between the corner points back to a distorted domain based on the camera parameters, and adjusting a direction of travel of the vehicle based on the located target object.

Abstract: Estimation of the ground plane of a three dimensional (3D) point cloud based modifications to the random sample consensus (RANSAC) algorithm is provided. The modifications may include applying roll and pitch constraints to the selection of random planes in the 3D point cloud, using a cost function based on the number of inliers in the random plane and the number of 3D points below the random plane in the 3D point cloud, and computing a distance threshold for the 3D point cloud that is used in determining whether or not a 3D point in the 3D point cloud is an inlier of a random plane.

Abstract: A method for video object detection includes detecting an object in a first video frame, and selecting a first interest point and a second interest point of the object. The first interest point is in a first region of interest located at a first corner of a box surrounding the object. The second interest point is in a second region of interest located at a second corner of the box. The second corner is diagonally opposite the first corner. A first optical flow of the first interest point and a second optical flow of the second interest point are determined. A location of the object in a second video frame is estimated by determining, in the second video frame, a location of the first interest point based on the first optical flow and a location of the second interest point based on the second optical flow.

Abstract: Joint denoising and supersampling of graphics data is described. An example of a graphics processor includes multiple processing resources, including a least a first processing resource including a pipeline to perform a supersampling operation; and the pipeline including circuitry to jointly perform denoising and supersampling of received ray tracing input data, the circuitry including first circuitry to receive input data associated with an input block for a neural network, second circuitry to perform operations associated with a feature extraction and kernel prediction network of the neural network, and third circuitry to perform operations associated with a filtering block of the neural network.

Abstract: A technique for key-point detection, including receiving, by a machine learning model, an input image, generating a set of image features for the input image, determining, by the machine learning model, based on the set of image features, a bounding box for an object detected in the input image, the bounding box described by bounding box information, identifying, by the machine learning model, based on the set of image features and a center point of the bounding box, a plurality of key-points associated with the object, filtering the plurality of key-points based on a confidence score associated with each key-point of the plurality of key-points, and outputting coordinates of the plurality of key-points, confidence scores associated with the plurality of key-points, and the bounding box information.

Abstract: A method for estimating time to collision (TTC) of a detected object in a computer vision system is provided that includes determining a three dimensional (3D) position of a camera in the computer vision system, determining a 3D position of the detected object based on a 2D position of the detected object in an image captured by the camera and an estimated ground plane corresponding to the image, computing a relative 3D position of the camera, a velocity of the relative 3D position, and an acceleration of the relative 3D position based on the 3D position of the camera and the 3D position of the detected object, wherein the relative 3D position of the camera is relative to the 3D position of the detected object, and computing the TTC of the detected object based on the relative 3D position, the velocity, and the acceleration.

Abstract: Disclosed examples include image processing methods and systems to process image data, including computing a plurality of scaled images according to input image data for a current image frame, computing feature vectors for locations of the individual scaled images, classifying the feature vectors to determine sets of detection windows, and grouping detection windows to identify objects in the current frame, where the grouping includes determining first clusters of the detection windows using non-maxima suppression grouping processing, determining positions and scores of second clusters using mean shift clustering according to the first clusters, and determining final clusters representing identified objects in the current image frame using non-maxima suppression grouping of the second clusters.

Abstract: One embodiment provides a graphics processor comprising a set of processing resources configured to perform a supersampling operation via a mixed precision convolutional neural network, the set of processing resources including circuitry configured to receive, at an input block of a neural network model, history data, velocity data, and current frame data, pre-process the history data, velocity data, and current frame data to generate pre-processed data, provide the pre-processed data to a feature extraction network of the neural network model, process the pre-processed data at the feature extraction network via one or more encoder stages and one or more decoder stages, and generate an output image via an output block of the neural network model via direct reconstruction or kernel prediction.

Abstract: A method for dynamically quantizing feature maps of a received image. The method includes convolving an image based on a predicted maximum value, a predicted minimum value, trained kernel weights and the image data. The input data is quantized based on the predicted minimum value and predicted maximum value. The output of the convolution is computed into an accumulator and re-quantized. The re-quantized value is output to an external memory. The predicted min value and the predicted max value are computed based on the previous max values and min values with a weighted average or a pre-determined formula. Initial min value and max value are computed based on known quantization methods and utilized for initializing the predicted min value and predicted max value in the quantization process.

Abstract: This invention predicts that intra mode prediction is more effective for the macroblocks where motion estimation in inter mode prediction fails. This failure is indicated by a large value of the inter mode SAD. This invention performs intra mode prediction for only macro blocks have larger inter mode SADs. The definition of a large inter mode SAD differs for different content. This invention compares the inter mode SAD of a current macroblock with an adaptive threshold. This adaptive threshold depends on the average and variance of the SADs of the previous predicted frame. An adaptive threshold is calculated for each new predictive frame.

Abstract: A method for generating a sparsified convolutional neural network (CNN) is provided that includes training the CNN to generate coefficient values of filters of convolution layers, and performing sparsified fine tuning on the convolution layers to generate the sparsified CNN, wherein the sparsified fine tuning causes selected nonzero coefficient values of the filters to be set to zero.

Abstract: A method for multi-label image classification in a convolutional neural network (CNN) is provided that includes forming a composite image from a plurality of clipped images, and processing the composite image by the CNN to generate a probability vector for each clipped image of the plurality of clipped images, wherein a length of a probability vector is equal to a number of classes the CNN is designed to classify.