Abstract: Videos captured in low light conditions can be processed in order to identify an activity being performed in the video. The processing may use both the video and audio streams for identifying the activity in the low light video. The video portion is processed to generate a darkness-aware feature which may be used to modulate the features generated from the audio and video features. The audio features may be used to generate a video attention feature and the video features may be used to generate an audio attention feature. The audio and video attention features may also be used in modulating the audio video features. The modulated audio and video features may be used to predict an activity occurring in the video.

Abstract: Repetitive activities can be captured in audio video content. The AV content can be processed in order to predict the number of repetitive activities present in the AV content. The accuracy of the predicted number may be improved, especially for AV content with challenging conditions, by basing the predictions on both the audio and video portions of the AV content.

Abstract: A transformer based vision-linguistic (VL) model and training technique uses a number of different image patches covering the same portion of an image, along with a text description of the image to train the model. The model and pre-training techniques may be used in domain specific training of the model. The model can be used for fine-grained image-text tasks in the fashion domain.

Abstract: This disclosure relates to improved sketch-based image retrieval (SBIR) techniques. The SBIR techniques utilize a neural network architecture to train a domain migration function and a hashing function. The domain migration function is configured to transform sketches into synthetic images, and the hashing function is configured to generate hash codes from synthetic images and authentic images in a manner that preserves semantic consistency across the sketch and image domains. The hash codes generated from the synthetic images can be used for accurately identifying and retrieving authentic images corresponding to sketch queries, or vice versa.

Abstract: Methods, systems, and techniques for kernel continuing learning. A dataset is obtained that corresponds to a classification task. Feature extraction is performed on the dataset using an artificial neural network. A kernel is constructed using features extracted during that feature extraction for use in performing the classification task. More particularly, during training, a coreset dataset corresponding to the classification task is saved; and during subsequent inference, the coreset dataset is retrieved and used to construct a task-specific kernel for classification.

Abstract: This disclosure relates to improved techniques for performing human parsing functions using neural network architectures. The neural network architecture can model human objects in images using a hierarchal graph of interconnected nodes that correspond to anatomical features at various levels. Multi-level inference information can be generated for each of the nodes using separate inference processes. The multi-level inference information for each node can be combined or fused to generate final predictions for each of the nodes. Parsing results may be generated based on the final predictions.

Abstract: This disclosure relates to improved techniques for performing computer vision functions, including common object detection and instance segmentation. The techniques described herein utilize neural network architectures to perform these functions in various types of images, such as natural images, UAV images, satellite images, and other images. The neural network architecture can include a dense location regression network that performs object localization and segmentation functions, at least in part, by generating offset information for multiple sub-regions of candidate object proposals, and utilizing this dense offset information to derive final predictions for locations of target objects. The neural network architecture also can include a discriminative region-of-interest (RoI) pooling network that performs classification of the localized objects, at least in part, by sampling various sub-regions of candidate proposals and performing adaptive weighting to obtain discriminative features.

Abstract: This disclosure relates to improved techniques for generating images from raw image sensor data captured in low-light conditions without the use of flash photography. The techniques described herein utilize a neural network architecture to transform the raw image sensor data into well-exposed images. The neural network architecture can be trained using a multi-criterion loss function that jointly models both pixel-level and feature-level properties of the images. The images output by the neural network architecture can be provided to a contrast correction module that enhances the contrast of the images.

Abstract: This disclosure relates to improved techniques for performing multi-view image clustering. The techniques described herein utilize machine learning functions to optimize the image clustering process. Multi-view features are extracted from a collection of images. A machine learning function is configured to jointly learn a fused binary representation that combines the multi-view features and one or more binary cluster structures that can be used to partition the images. A clustering function utilizes the fused binary representation and the one or more binary cluster structures to generate one or more image clusters based on the collection of images.

Abstract: This disclosure relates to improved techniques for performing computer vision functions including motion deblurring functions. The techniques described herein utilize a neural network architecture to perform these functions. The neural network architecture can include a human-aware attention model that is able to distinguish between foreground human objects and background portions of degraded images affected by motion blur. The neural network architecture further includes an encoder-decoder network that separately performs motion deblurring functions on foreground and background portions of degraded images, and reconstructs enhanced images corresponding to the degraded images.

Abstract: This disclosure relates to improved techniques for performing computer vision functions including common object counting and instance segmentation. The techniques described herein utilize a neural network architecture to perform these functions. The neural network architecture can be trained using image-level supervision techniques that utilize a loss function to jointly train an image classification branch and a density branch of the neural network architecture. The neural network architecture constructs per-category density maps that can be used to generate analysis information comprising global object counts and locations of objects in images.

Abstract: This disclosure relates to improved techniques for performing computer vision functions on medical images, including object segmentation functions for identifying medical objects in the medical images and grading functions for determining severity labels for medical conditions exhibited in the medical images. The techniques described herein utilize a neural network architecture to perform these and other functions. The neural network architecture can be trained, at least in part, using semi-supervised learning techniques that enable the neural network architecture to accurately perform the object segmentation and grading functions despite limited availability of pixel-level annotation information.

Abstract: This disclosure relates to improved techniques for synthesizing three-dimensional (3D) scenes. The techniques can utilize a neural network architecture to analyze images for detecting objects, classifying scenes and objects, and determining degree of freedom information for objects in the images. These tasks can be performed by, at least in part, using inter-object and object-scene dependency information that captures the spatial correlations and dependencies among objects in the images, as well as the correlations and relationships of objects to scenes associated with the images. 3D scenes corresponding to the images can then be synthesized using the inferences provided by the neural network architecture.

Abstract: 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.

Abstract: This disclosure relates to improved vehicle re-identification techniques. The techniques described herein utilize artificial intelligence (AI) and machine learning functions to re-identify vehicles across multiple cameras. Vehicle re-identification can be performed using an image of the vehicle that is captured from any single viewpoint. Attention maps may be generated that identify regions of the vehicle that include visual patterns that overlap between the viewpoint of the captured image and one or more additional viewpoints. The attention maps are used to generate a multi-view representation of the vehicle that provides a global view of the vehicle across multiple viewpoints. The multi-view representation of the vehicle can then be compared to previously captured image data to perform vehicle re-identification.