Abstract: Certain aspects of the present disclosure provide techniques for lane marker detection. A set of feature tensors is generated by processing an input image using a convolutional neural network. A set of localizations is generated by processing the set of feature tensors using a localization network, a set of horizontal positions is generated by processing the set of feature tensors using row-wise regression, and a set of end positions is generated by processing the set of feature tensors using y-end regression. A set of lane marker positions is determined based on the set of localizations, the set of horizontal positions, and the set of end positions.

Abstract: A processor-implemented method of searching for a neural network architecture includes defining a search space of student neural network architectures for knowledge distillation. The search space includes multiple convolutional operators and multiple transformer operators. A trust-region Bayesian optimization is performed to select a student neural network architecture from the search space based on a pre-defined teacher model.

Abstract: Certain aspects of the present disclosure provide a method for lane marker detection, including: receiving an input image; providing the input image to a lane marker detection model; processing the input image with a shared lane marker portion of the lane marker detection model; processing output of the shared lane marker portion of the lane marker detection model with a plurality of lane marker-specific representation layers of the lane marker detection model to generate a plurality of lane marker representations; and outputting a plurality of lane markers based on the plurality of lane marker representations.

Abstract: Disclosed are techniques for performing lane instance recognition. Lane instances are difficult to recognize since they are long and elongated, and they also look different from view to view. An approach is proposed in which local mask segmentation lane estimation and global control points lane estimation are combined.

Abstract: Certain aspects of the present disclosure are directed to methods and apparatus for deep learning in an artificial neural network. One example method generally includes receiving input data at an input to a layer of the neural network; replicating a group of neural processing units in the layer to form a superset of neural processing units, the superset comprising n instances of the group of neural processing units; processing the input data using the superset to generate output data for the layer; and determining an uncertainty of the output data. Processing the input data includes performing a dropout function by zeroing out one or more weights of a set of weights for each of the n instances of the superset of neural processing units and convolving, for each of the n instances in parallel, the input data with one or more non-zeroed out weights of the set of weights.

Abstract: Certain aspects of the present disclosure provide a method for lane marker detection, including: receiving an input image; providing the input image to a lane marker detection model; processing the input image with a shared lane marker portion of the lane marker detection model; processing output of the shared lane marker portion of the lane marker detection model with a plurality of lane marker-specific representation layers of the lane marker detection model to generate a plurality of lane marker representations; and outputting a plurality of lane markers based on the plurality of lane marker representations.

Abstract: Autonomous driving systems described herein provide an efficient way to manage camera-based perception by considering the characteristics of captured images. In one example, a camera sensor may capture an image and a processor may determine a first region of interest (ROI) within the image and a second ROI within the image. The processor may generate a first image of the first ROI and a second image of the second ROI. The processor may transmit a control signal based on one or more objects detected in the first ROI and/or one or more objects detected in the second ROI to cause the vehicle to perform an autonomous driving operation.

Abstract: Disclosed are techniques for performing lane instance recognition. Lane instances are difficult to recognize since they are long and elongated, and they also look different from view to view. An approach is proposed in which local mask segmentation lane estimation and global control points lane estimation are combined.

Abstract: Certain aspects of the present disclosure are directed to methods and apparatus for deep learning in an artificial neural network. One example method generally includes receiving input data at an input to a layer of the neural network; replicating a group of neural processing units in the layer to form a superset of neural processing units, the superset comprising n instances of the group of neural processing units; processing the input data using the superset to generate output data for the layer; and determining an uncertainty of the output data. Processing the input data includes performing a dropout function by zeroing out one or more weights of a set of weights for each of the n instances of the superset of neural processing units and convolving, for each of the n instances in parallel, the input data with one or more non-zeroed out weights of the set of weights.