A method in a vehicle for performing multiple on-board sensing tasks concurrently in the same network using deep learning algorithms is provided. The method includes receiving vision sensor data from a sensor on the vehicle, determining a set of features from the vision sensor data using a plurality of feature layers in a convolutional neural network, and concurrently estimating, using the convolutional neural network, bounding boxes for detected objects, free-space boundaries, and object poses for detected objects from the set of features determined by the plurality of feature layers. The neural network may include a plurality of free-space estimation layers configured to determine the boundaries of free-space in the vision sensor data, a plurality of object detection layers configured to detect objects in the image and to estimate bounding boxes that surround the detected objects, and a plurality of object pose detection layers configured to estimate the direction of each object.

Unified deep convolutional neural net for free-space estimation, object detection and object pose estimation

Methods, systems, and apparatus, including computer programs encoded on computer storage medium, for regularizing feature weights maintained by a machine learning model. The method includes actions of obtaining a set of training data that includes multiple training feature vectors, and training the machine learning model on each of the training feature vectors, comprising, for each feature vector and for each of a plurality of the features of the feature vector: determining a first loss for the feature vector with the feature, determining a second loss for the feature vector without the feature, and updating a current benefit score for the feature using the first loss and the second loss, wherein the benefit score for the feature is indicative of the usefulness of the feature in generating accurate predicted outcomes for training feature vectors.

Regularization of machine learning models

One embodiment provides a method comprising receiving a first input image and a second input image. Each input image comprises a facial image of an individual. For each input image, a first set of facial regions of the facial image is distinguished from a second set of facial regions of the facial image based on a learning based model. The first set of facial regions comprises age-invariant facial features, and the second set of facial regions comprises age-sensitive facial features. The method further comprises determining whether the first input image and the second input images comprise facial images of the same individual by performing face verification based on the first set of facial regions of each input image.

Learning model for salient facial region detection

This disclosure relates to improved sketch-based image retrieval (SBIR) techniques. The SBIR techniques utilize an architecture comprising three interconnected neural networks to enable zero-shot image recognition and retrieval based on free-hand sketches. Zero-shot learning may be implemented to retrieve one or more images corresponding to the sketches without prior training on all categories of the sketches. The neural network architecture may do so, at least in part, by training encoder hashing functions to mitigate heterogeneity of sketches and images, and by applying semantic knowledge that is learned during a limited training phase to unknown categories.

Zero-shot sketch-based image retrieval techniques using neural networks for sketch-image recognition and retrieval

One embodiment provides a method comprising estimating a camera pose of an input image and aligning the input image to a desired camera pose based on a feature database. The input image comprises an image of a fine-grained object. The method further comprises classifying the object based on the alignment.

Leveraging multi cues for fine-grained object classification

A method includes receiving, at a neural network, a subset of images of a plurality of images of a training image set. The method includes training the neural network by iteratively adjusting parameters of the neural network based on concurrent application of multiple loss functions to the subset of images. The multiple loss functions include a classification loss function and a hashing loss function. The classification loss function is associated with an image classification function that extracts image features from an image. The hashing loss function is associated with a hashing function that generates a hash code for the image.

Hash codes for images

A method for classifying images using a neural network is described. The method utilizes the neural network to generate a deep hash data structure. In one embodiment, the method includes receiving, at a neural network, first data corresponding to a plurality of images of a training image set. The method includes adjusting parameters of the neural network based on concurrent application at the neural network of multiple loss functions to second data corresponding to the plurality of images to generate adjusted parameters. The method includes generating a deep hash data structure based on the adjusted parameters of the neural network. The method further includes sending, via a transmitter to a device, third data corresponding to the deep hash data structure.

Yadong Mu

Papers

Hash codes for images

Image hash codes generated by a neural network