Adobe, Inc.

Apple Inc.

Kshitiz Garg

Ali Aminian

Ahmad Al-Dahle

Aseem Agarwala

Akhilesh Kumar

Daniel Miranda

Aashish Misraa

Xiaozhen Xue

Han Guo

Viswanathan Swaminathan

Ryan A Rossi

Saayan Mitra

Randol W Aikin

Somdeb Sarkhel

Xiang Chen

Nicolas Huynh Thien

Xinyu Xu

YoungWoo Seo

Shivam Nalin Patel

Swapneel Mehta

Embodiments of the present invention provide systems, methods, and computer storage media for extracting moments of interest (e.g., video frames, video segments) from a video. In an example embodiment, independent and/or orthogonal machine learning models are used to extract different types of features considering different modalities, and each frame in the video is assigned an importance score for each model. The importance scores for each model are combined into an aggregated importance score for each frame in the video. Depending on the embodiment, the aggregated importance scores are used to visualize the score per frame, identify moments of interest, automatically crop down the video into a highlight reel, browse or visualize the moments of interest within the video, and/or search across multiple videos.

Customizable framework to extract moments of interest

This disclosure describes one or more implementations of a video inference system that utilizes machine-learning models to efficiently and flexibly process digital videos utilizing various improved video inference architectures. For example, the video inference system provides a framework for improving digital video processing by increasing the efficiency of both central processing units (CPUs) and graphics processing units (GPUs). In one example, the video inference system utilizes a first video inference architecture to reduce the number of computing resources needed to inference digital videos by analyzing multiple digital videos utilizing sets of CPU/GPU containers along with parallel pipeline processing. In a further example, the video inference system utilizes a second video inference architecture that facilitates multiple CPUs to preprocess multiple digital videos in parallel as well as a GPU to continuously, sequentially, and efficiently inference each of the digital videos.

Efficiently inferencing digital videos utilizing machine-learning models

Some aspects of the technology described herein perform identity identification on faces in a video. Object tracking is performed on detected faces in frames of a video to generate tracklets. Each tracklet comprises a sequence of consecutive frames in which each frame includes a detected face for a person. The tracklets are clustered using face feature vectors for detected faces of each tracklet to generate a plurality of clusters. Information is stored in an identity datastore, including a first identifier for a first identity in association with an indication of frames from tracklets in a first cluster from the plurality of clusters.

Tracking unique face identities in videos

Embodiments are disclosed for receiving a user input and an input video comprising multiple frames. The method may include extracting a text feature from the user input. The method may further include extracting a plurality of image features from the frames. The method may further include identifying one or more keyframes from the frames that include the object. The method may further include clustering one or more groups of the one or more keyframes. The method may further include generating a plurality of segmentation masks for each group. The method may further include determining a set of reference masks corresponding to the user input and the object. The method may further include generating a set of fusion masks by combining the plurality of segmentation masks and the set of reference masks. The method may further include propagating the set of fusion masks and outputting a final set of masks.

Text-based framework for video object selection

Increasing efficiency of inferencing digital videos utilizing machine-learning models

Techniques for recommending hashtags, including trending hashtags, are disclosed. An example method includes accessing a graph. The graph includes video nodes representing videos, historical hashtag nodes representing historical hashtags, and edges indicating associations among the video nodes and the historical hashtag nodes. A trending hashtag is identified. An edge is added to the graph between a historical hashtag node representing a historical hashtag and a trending hashtag node representing the trending hashtag, based on a semantic similarity between the historical hashtag and the trending hashtag. A new video node representing a new video is added to the video nodes of the graph. A graph neural network (GNN) is applied to the graph, and the GNN predicts a new edge between the trending hashtag node and the new video node. The trending hashtag is recommended for the new video based on prediction of the new edge.

Open-domain trending hashtag recommendations

Some embodiments provide a sensor data-processing system which detects and classifies objects detected in an environment via fusion of sensor data representations generated by multiple separate sensors. The sensor data-processing system can fuse sensor data representations generated by multiple sensor devices into a fused sensor data representation and can further detect and classify features in the fused sensor data representation. Feature detection can be implemented based at least in part upon utilizing a feature-detection model generated via one or more of deep learning and traditional machine learning. The sensor data-processing system can adjust sensor data processing of representations generated by sensor devices based on external factors including indications of sensor health and environmental conditions. The sensor data-processing system can be implemented in a vehicle and provide output data associated with the detected objects to a navigation system which navigates the vehicle according to the output data.

Sensor fusion and deep learning

Sensor data captured at by different sensors may be shared across different sensor processing pipelines. Sensor processing pipelines may process captured sensor data from respective sensors. Some of the sensor data that is received or processed at one sensor data processing pipeline may be provided to another sensor data processing pipeline so that subsequent processing stages at the recipient sensor processing pipeline may process the combined sensor data in order to determine a perception decision. Different types of sensor data may be shared, including raw sensor data, processed sensor data, or data derived from sensor data. A control system may perform control actions based on the perception decisions determined by the sensor processing pipelines that share sensor data.

Shared sensor data across sensor processing pipelines

Aspects of the present disclosure involve systems and methods for obtaining real-time road friction coefficient estimations. In one embodiment, a regression function is learned using a training data set which correlates input data measurements arriving from onboard system sensors and coefficient estimations arriving from an extension system. In another embodiment, the learned regression function can be retrieved to obtain real-time road friction coefficient estimations while the system is in motion.

Growing community of inventors

Santa Clara, CA, United States of America

Kshitiz Garg