L'oreal

Modiface Inc.

Other

De-identification Ltd.

Loreal Sa

Elmoznino, Eric

Flament, Frederic Antoinin Raymond Serge

Jiang, Ruowei

Kezele, Irina

Ma, Junwei

Levinshtein, Alex

Despois, Julien

Aarabi, Parham

Ma, He

Perrot, Matthieu

Parham Aarabi

Irina Kezele

Ruowei Jiang

Edmund Phung

Eric Elmoznino

Alex Levinshtein

Zhi Yu

Frederic Antoinin Raymond Serge Flament

Brendan Duke

Julien Despois

Yuze Zhang

He Ma

Matthieu Perrot

Cheng Chang

Junwei Ma

Tian Xing Li

Abdalla Ahmed

Zeqi Li

Angel Jing Yi Zhang

Wenzhangzhi Guo

Ron Appel

Zihao Chen

Jeffrey Houghton

Avishek Bose

Sophie Seite

Mathieu Perrot

Cong Wei

Tian Yu Tommy Liu

Vision Transformers (ViT) have shown their competitive advantages performance-wise compared to convolutional neural networks (CNNs) though they often come with high computational costs. Methods, systems and techniques herein learn instance-dependent attention patterns, utilizing a lightweight connectivity predictor module to estimate a connectivity score of each pair of tokens. Intuitively, two tokens have high connectivity scores if the features are considered relevant either spatially or semantically. As each token only attends to a small number of other tokens, the binarized connectivity masks are often very sparse by nature providing an opportunity to accelerate the network via sparse computations. Equipped with the learned unstructured attention pattern, sparse attention ViT produces a superior Pareto-optimal trade-off between FLOPs and top-1 accuracy on ImageNet compared to token sparsity (48%˜69% FLOPs reduction of MHSA; accuracy drop within 0.4%). Combining attention and token sparsity reduces VIT FLOPs by over 60%.

Systems, methods and techniques for learning and using sparse instance-dependent attention for efficient vision transformers

Systems, methods and techniques provide for acne localization, counting and visualization. An image is processed using a trained model to identify objects. The model may be a deep learning (e.g. convolutional neural) network configured for object classification with a detection focus on small objects. The image may be a frontal or profile facial image, processed end to end. The model identifies and localizes different types of acne. Instances are counted and visualized such as by annotating the source image. An example annotation is an overlay identifying a type and location of each instance. Counts by acne type assist with scoring. A product and/or service may be recommended in response to the identification of the acne (e.g. the type, localization, counting and/or a score).

Systems and methods for acne counting, localization and visualization

There are provided systems and methods for facial landmark detection using a convolutional neural network (CNN). The CNN comprises a first stage and a second stage where the first stage produces initial heat maps for the landmarks and initial respective locations for the landmarks. The second stage processes the heat maps and performs Region of Interest-based pooling while preserving feature alignment to produce cropped features. Finally, the second stage predicts from the cropped features a respective refinement location offset to each respective initial location. Combining each respective initial location with its respective refinement location offset provides a respective final coordinate (x,y) for each respective landmark in the image. Two-stage localization design helps to achieve fine-level alignment while remaining computationally efficient. The resulting architecture is both small enough in size and inference time to be suitable for real-time web applications such as product simulation and virtual reality.

Convolution neural network based landmark tracker

With Convolutional Neural Networks (CNN), facial alignment networks (FAN) have achieved significant accuracy on a wide range of public datasets, which comes along with larger model size and expensive computation costs, making it infeasible to adapt them to real-time applications on edge devices. There is provided a model compression approach on FAN using One-Shot Neural Architecture Search to overcome this problem while preserving performance criteria. Methods and devices provide efficient training and searching (on a single GPU), and resultant models can deploy to run real-time in browser-based applications on edge devices including tablets and smartphones. The compressed models provide comparable cutting-edge accuracy, while having a 30 times smaller model size and can run 40.7 ms per frame in a popular browser on a popular smartphone and OS.

Lightweight real-time facial alignment network model selection process

There is provided methods, devices and techniques to process an image using a deep learning model to achieve continuous effect simulation by a unified network where a simple (effect class) estimator is embedded into a regular encoder-decoder architecture. The estimator allows learning of model-estimated class embeddings of all effect classes (e.g. progressive degrees of the effect), thus representing the continuous effect information without manual efforts in selecting proper anchor effect groups. In an embodiment, given a target age class, there is derived a personalized age embedding which considers two aspects of face aging: 1) a personalized residual age embedding at a model-estimated age of the subject, preserving the subject's aging information; and 2) exemplar-face aging basis at the target age, encoding the shared aging patterns among the entire population. Training and runtime (inference time) embodiments are described including an AR application that generates recommendations and provides ecommerce services.

Applying a continuous effect via model-estimated class embeddings

GANs based generators are useful to perform image to image translations. GANs models have large storage sizes and resource use requirements such that they are too large to be deployed directly on mobile devices. Systems and methods define through conditioning a student GANs model having a student generator that is scaled downwardly from a teacher GANs model (and generator) using knowledge distillation. A semantic relation knowledge distillation loss is used to transfer semantic knowledge from an intermediate layer of the teacher to an intermediate layer of the student. Student generators thus defined are stored and executed by mobile devices such as smartphones and laptops to provide augmented reality experiences. Effects are simulated on images, including makeup, hair, nail and age simulation effects.

Semantic relation preserving knowledge distillation for image-to-image translation

Techniques are provided for computing systems, methods and computer program products to produce efficient image-to-image translation by adapting unpaired datasets for supervised learning. A first model (a powerful model) may be defined and conditioned using unsupervised learning to produce a synthetic paired dataset from the unpaired dataset, translating images from a first domain to a second domain and images from the second domain to the first domain. The synthetic data generated is useful as ground truths in supervised learning. The first model may be conditioned to overfit the unpaired dataset to enhance the quality of the paired dataset (e.g. the synthetic data generated). A run-time model such as for a target device is trained using the synthetic paired dataset and supervised learning. The run-time model is small and fast to meet the processing resources of the target device (e.g. a personal user device such as a smart phone, tablet, etc.)

Image-to-image translation using unpaired data for supervised learning

There is described a deep learning supervised regression based model including methods and systems for facial attribute prediction and use thereof. An example of use is an augmented and/or virtual reality interface to provide a modified image responsive to facial attribute predictions determined from the image. Facial effects matching facial attributes are selected to be applied in the interface.

Systems and methods for improved facial attribute classification and use thereof

Systems and methods process images to determine a skin condition severity analysis and to visualize a skin analysis such as using a deep neural network (e.g. a convolutional neural network) where a problem was formulated as a regression task with integer-only labels. Auxiliary classification tasks (for example, comprising gender and ethnicity predictions) are introduced to improve performance. Scoring and other image processing techniques may be used (e.g. in assoc. with the model) to visualize results such as highlighting the analyzed image. It is demonstrated that the visualization of results, which highlight skin condition affected areas, can also provide perspicuous explanations for the model. A plurality (k) of data augmentations may be made to a source image to yield k augmented images for processing. Activation masks (e.g. heatmaps) produced from processing the k augmented images are used to define a final map to visualize the skin analysis.

Systems and methods to process images for skin analysis and to visualize skin analysis

A system and method implement deep learning on a mobile device to provide a convolutional neural network (CNN) for real time processing of video, for example, to color hair. Images are processed using the CNN to define a respective hair matte of hair pixels. The respective object mattes may be used to determine which pixels to adjust when adjusting pixel values such as to change color, lighting, texture, etc. The CNN may comprise a (pre-trained) network for image classification adapted to produce the segmentation mask. The CNN may be trained for image segmentation (e.g. using coarse segmentation data) to minimize a mask-image gradient consistency loss. The CNN may further use skip connections between corresponding layers of an encoder stage and a decoder stage where shallower layers in the encoder, which contain high-res but weak features are combined with low resolution but powerful features from deeper decoder layers.

System and method for image processing using deep neural networks

There is shown and described a deep learning based system and method for skin diagnostics as well as testing metrics that show that such a deep learning based system outperforms human experts on the task of apparent skin diagnostics. Also shown and described is a system and method of monitoring a skin treatment regime using a deep learning based system and method for skin diagnostics.

Automatic image-based skin diagnostics using deep learning

This document relates to hybrid eye center localization using machine learning, namely cascaded regression and hand-crafted model fitting to improve a computer. There are proposed systems and methods of eye center (iris) detection using a cascade regressor (cascade of regression forests) as well as systems and methods for training a cascaded regressor. For detection, the eyes are detected using a facial feature alignment method. The robustness of localization is improved by using both advanced features and powerful regression machinery. Localization is made more accurate by adding a robust circle fitting post-processing step. Finally, using a simple hand-crafted method for eye center localization, there is provided a method to train the cascaded regressor without the need for manually annotated training data. Evaluation of the approach shows that it achieves state-of-the-art performance.

System and method using machine learning for iris tracking, measurement, and simulation

There are provided methods and computing devices using semi-supervised learning to perform end-to-end video object segmentation, tracking respective object(s) from a single-frame annotation of a reference frame through a video sequence of frames. A known deep learning model may be used to annotate the reference frame to provide ground truth locations and masks for each respective object. A current frame is processed to determine current frame object locations, defining object scoremaps as a normalized cross-correlation between encoded object features of the current frame and encoded object features of a previous frame. Scoremaps for each of more than one previous frame may be defined. An Intersection over Union (IoU) function, responsive to the scoremaps, ranks candidate object proposals defined from the reference frame annotation to associate the respective objects to respective locations in the current frame. Pixel-wise overlap may be removed using a merge function responsive to the scoremaps.

End-to-end merge for video object segmentation (VOS)

Systems and methods relate to a network model to apply an effect to an image such as an augmented reality effect (e.g. makeup, hair, nail, etc.). The network model uses a conditional cycle-consistent generative image-to-image translation model to translate images from a first domain space where the effect is not applied and to a second continuous domain space where the effect is applied. In order to render arbitrary effects (e.g. lipsticks) not seen at training time, the effect's space is represented as a continuous domain (e.g. a conditional variable vector) learned by encoding simple swatch images of the effect, such as are available as product swatches, as well as a null effect. The model is trained end-to-end in an unsupervised fashion. To condition a generator of the model, convolutional conditional batch normalization (CCBN) is used to apply the vector encoding the reference swatch images that represent the makeup properties.

System and method for augmented reality using conditional cycle-consistent generative image-to-image translation models

Growing community of inventors

Richmond Hill, Canada

Parham Aarabi