Snap Inc.

Lemon Inc.

Linjie Yang

Jianchao Yang

Xuehan Xiong

Samuel Edward Hare

Fedir Poliakov

Guohui Wang

William Brendel

Shah Tanmay Anilkumar

Lidiia Bogdanovych

Jia Li

Kevin Dechau Tang

Ding Liu

Yanran Wang

Ning Zhang

Yuncheng Li

Xiaojie Jin

Zhengyuan Yang

Yiming Cui

Peng Wang

Longyin Wen

Xiaochen Lian

Xueyan Zou

Taojiannan Yang

Peter Lin

Imran Saleemi

Mingyu Ding

Haichao Yu

Ziyu Jiang

Object detection using a transformer-based object detection model includes randomly initializing basic queries for the model, modulating the basic queries based on semantics of input images, and training the model basic on features extracted from input images and the modulated queries.

Transformer-based object detection

The present disclosure describes techniques of improving video matting. The techniques comprise extracting features from each frame of a video by an encoder of a model, wherein the video comprises a plurality of frames; incorporating, by a decoder of the model, into any particular frame temporal information extracted from one or more frames previous to the particular frame, wherein the particular frame and the one or more previous frames are among the plurality of frames of the video, and the decoder is a recurrent decoder; and generating a representation of a foreground object included in the particular frame by the model, wherein the model is trained using segmentation dataset and matting dataset.

Video matting

Described are examples for detecting objects in an image on a device including setting, based on a condition, a number of sparse proposals to use in performing object detection in the image, performing object detection in the image based on providing the sparse proposals as input to an object detection process to infer object location and classification of one or more objects in the image, and indicating, to an application and based on an output of the object detection process, the object location and classification of the one or more objects.

Techniques for using dynamic proposals in object detection

A modulated segmentation system can use a modulator network to emphasize spatial prior data of an object to track the object across multiple images. The modulated segmentation system can use a segmentation network that receives spatial prior data as intermediate data that improves segmentation accuracy. The segmentation network can further receive visual guide information from a visual guide network to increase tracking accuracy via segmentation.

Modulated image segmentation

A machine learning system can generate an image mask (e.g., a pixel mask) comprising pixel assignments for pixels. The pixels can be assigned to classes, including, for example, face, clothes, body skin, or hair. The machine learning system can be implemented using a convolutional neural network that is configured to execute efficiently on computing devices having limited resources, such as mobile phones. The pixel mask can be used to more accurately display video effects interacting with a user or subject depicted in the image.

Generating an image mask using machine learning

A mobile device can generate real-time complex visual image effects using asynchronous processing pipeline. A first pipeline applies a complex image process, such as a neural network, to keyframes of a live image sequence. A second pipeline generates flow maps that describe feature transformations in the image sequence. The flow maps can be used to process non-keyframes on the fly. The processed keyframes and non-keyframes can be used to display a complex visual effect on the mobile device in real-time or near real-time.

Real-time tracking-compensated image effects

Systems and methods for obtaining attention features are described. Some examples may include: receiving, at a projector of a transformer, a plurality of tokens associated with image features of a first dimensional space; generating, at the projector of the transformer, projected features by concatenating the plurality of tokens with a positional map, the projected features having a second dimensional space that is less than the first dimensional space; receiving, at an encoder of the transformer, the projected features and generating encoded representations of the projected features using self-attention; decoding, at a decoder of the transformer, the encoded representations and obtaining a decoded output; and projecting the decoded output to the first dimensional space and adding the image features of the first dimensional space to obtain attention features associated with the image features.

Lightweight transformer for high resolution images

Systems and methods for performing neural architecture search are provided. In one aspect, the system includes a processor configured to select a plurality of candidate neural networks within a search space, evaluate a performance of each of the plurality of candidate neural networks by: training each candidate neural network on a training dataset to perform the predetermined task and determining a ranking metric for each candidate neural network based on an objective function. The ranking metric includes a weight-related metric that is determined based on weights of a prediction layer of each respective candidate neural network before and after the respective candidate neural network is trained. The processor is configured to rank the plurality of candidate neural networks based on the determined ranking metrics.

Neural architecture search system using training based on a weight-related metric

A machine learning system can generate an image mask (e.g., a pixel mask) comprising pixel assignments for pixels. The pixels can he assigned to classes, including, for example, face, clothes, body skin, or hair. The machine learning system can be implemented. using a convolutional neural network that is configured to execute efficiently on computing devices having limited resources, such as mobile phones. The pixel mask can be used to more accurately display video effects interacting with a user or subject depicted in the image.

Segmentation of an image into individual body parts is performed based on a trained model. The model is trained with a plurality of training images, each training image representing a corresponding training figure. The model is also trained with a corresponding plurality of segmentations of the training figures. Each segmentation is generated by positioning body parts between defined positions of joints of the represented figure. The body parts are represented by body part templates obtained from a template library, with the templates defining characteristics of body parts represented by the templates.

Weakly supervised semantic parsing

Systems and methods are directed to inpainting video. More specifically, initial video data including a sequence of image frames containing missing or corrupted pixel information may be received. Optical flow displacement values and optical flow validity masks may be generated for neighboring image frames of initial video data. Image features from image feature maps of one or more neighboring image frames may be warp-shifted to image feature maps of a current image frame using the optical flow displacement values and warp-shifted image features from the feature maps of the one or more neighboring image frames may be selected based on one or more of the optical flow validity masks. A sequence of complete image frames may be generated based on the selected warp-shifted image features from the feature maps of the one or more neighboring image frames and image features from the image feature maps of the current image frame.

Temporal feature alignment network for video inpainting

Systems and method directed to performing video object segmentation are provided. In examples, video data representing a sequence of image frames and video data representing an object mask may be received at a video object segmentation server. Image features may be generated based on a first image frame of the sequence of image frames, image features may be generated based on a second image frame of the sequence of image frames; and object features may be generated based on the object mask. A transform matrix may be computed based on the image features of the first image frame and image features of the second image frame; the transform matrix may be applied to the object features resulting in transformed object features. A predicted object mask associated with the second image frame may be obtained by decoding the transformed object features.

Growing community of inventors

Los Angeles, CA, United States of America

Linjie Yang