Lemon Inc.

Shuo Cheng

Linjie Luo

Wanchun Ma

Chao Wang

Michael Leong Hou Tay

Peng Wang

Systems and methods for multi-task joint training of a neural network including an encoder module and a multi-headed attention mechanism are provided. In one aspect, the system includes a processor configured to receive input data including a first set of labels and a second set of labels. Using the encoder module, features are extracted from the input data. Using a multi-headed attention mechanism, training loss metrics are computed. A first training loss metric is computed using the extracted features and the first set of labels, and a second training loss metric is computed using the extracted features and the second set of labels. A first mask is applied to filter the first training loss metric, and a second mask is applied to filter the second training loss metric. A final training loss metric is computed based on the filtered first and second training loss metrics.

Systems for multi-task joint training of neural networks using multi-label datasets

The present disclosure describes techniques for determining a bounding box. An image may be received. An X-frame, a Y-frame, and a normal frame may be estimated based on the image using a first neural network. At least one planar region may be detected from the image using a second neural network. A vanishing point detection may be performed on each of the at least one planar region. Output of the first neural network may be fused with results of the vanishing point detection. A depth value of each pixel in at least one plane corresponding to the at least one planar region may be determined based at least in part on a result of the fusing. A location of a bounding box may be determined based at least in part on the depth value of each pixel in the at least one plane.

Box detection for object attachment

The present disclosure describes techniques for facial expression recognition. A first loss function may be determined based on a first set of feature vectors associated with a first set of images depicting facial expressions and a first set of labels indicative of the facial expressions. A second loss function may be determined based on a second set of feature vectors associated with a second set of images depicting asymmetric facial expressions and a second set of labels indicative of the asymmetric facial expressions. The first loss function and the second loss function may be used to determine a maximum loss function. The maximum loss function may be applied during training of a model. The trained model may be configured to predict at least one asymmetric facial expression in a subsequently received image.

Asymmetric facial expression recognition

Techniques for face tracking comprise receiving landmark data associated with a plurality of images indicative of at least one facial part. Representative images corresponding to the plurality of images may be generated based on the landmark data. Each representative image may depict a plurality of segments, and each segment may correspond to a region of the at least one facial part. The plurality of images and corresponding representative images may be input into a neural network to train the neural network to predict a feature associated with a subsequently received image comprising a face. An animation associated with a facial expression may be controlled based on output from the trained neural network.

Growing community of inventors

Los Angeles, CA, United States of America

Shuo Cheng