Meeting Deployment Needs on Cloud / Mobile / Edge Devices
With Brain++ serving as the framework, we have developed state-of-the-art neural networks that can be deployed to train algorithms on different computing platforms: cloud centers, mobile devices and edge servers. Which helped our full-stack solutions optimize the distribution of computing power and data storage, thereby improving performance and efficiency.
Cloud-based Deep Neural Network ResNet (advanced)
- ResNet (advanced) has broken the depth limit of network models, and is a milestone in deep learning area
- It has been widely used by the industry, and has become the basic model of the AI industry
- It is suitable for cloud visual computing and processing
Edge-based Neural Network DoReFa-Net
- The first model that can simultaneously quantify weights, activations, and gradients
- DoReFa-Net may accelerate the development and improvement of cloud and terminal AI chips
- It is suitable for CPU, GPU, FPGA and ASIC platforms
Mobile-based Neural Network shuffleNet
- A more efficient lightweight convolutional neural network which has opened its source on Github
- ShuffleNet has achieved high-precision identification while dramatically reducing the complexity of model calculation
- It can be used in mobile portrait retouching, image beautification, interactive live webcast and other products
Covering Mainstream AI Application Scenarios
In specific applications, MEGVII classifies deep learning algorithms into three categories: sensing, actuating and optimization. The company is dedicated to achieving cognition and decision-making from human faces, the environment, and various other objects. These core algorithms have been applied and implemented in MEGVII’s business scenarios in varying degrees.
Detect and locate subjects including human faces,human bodies, vehicles and various other general objects within images and videos and return high-precision object bounding boxes.
Check the likelihood that two subjects belong to the same person or that two objects are the same and generate a confidence score to measure the similarity.
Dense Facial Landmarks
Accurately locate and generate hundreds of landmarks of facial features, including face contour, eye, eyebrow, lip and nose contour.
Analyze a series of face related attributes including age, gender, smile intensity, head pose, eye status, emotion, beauty, eye gaze, mouth status, skin status, image quality and blurriness.
Analyze and identify emotion of detected faces. Analysis result of each detected face includes confidence scores for several kinds of emotions.
Object Instances/Scene Segmentation
Separate object instances or semantic regions from the background within images and videos and return high-precision object/region boundaries.
Detect and recognize hand/body gestures and human actions and generate high-precision confidence score.
Large Scale Photo Clustering and Search
Recognize objects in a large scale photo collection and automatically group the photos of the same object or instance together for easy photo search.
Control and Optimization Technologies in Robotics
Depth and Landmark Sensing
This technology uses a pair of cameras or a pair of camera and projector or a single camera, to compute depth or locate landmarks of environment where the robot is, or the object to be manipulated.The depth or landmarks is computed by a deep learning method which directly learns the sensing capabilities from a large amount of training examples.
Deep SLAM (Simultaneous Localization and Mapping)
This technology is the process of creating a map using a robot that navigates that environment while using the map it generates. In the localization, deep learning global visual features are used for robustness and deep learning local visual features are applied for accuracy. In the mapping, the whole system integrates multiple inputs to construct the map in real-time.
6D Object Detection and Pose Estimation
This technology detects and estimates 6D pose (3D location + 3D orientation) of objects to be gripped or manipulated. Taking both camera input and depth sensing, accurate object pose is obtained by an end-to-end learned estimator which is offline trained on various objects and pose variations.
This technology finds a sequence of movements or valid configurations that moves a robot or a robotic arm from the source state to destination state. In both planning tasks, we use an efficient sampling-based approach to obtain a fast and reliable initial solution and keep improve the planning strategy by a reinforcement learning approach which can learn from experiences and improve performance over time.
Collaborative Optimization Technologies in Robotics
Multi-robot Traffic Scheduling
This technology provides a unified traffic command system, allowing multiple mobile robots to move in an orderly manner in a unified area, avoiding congestion and optimizing the traffic efficiency of the entire robot group.
Job Scheduling & Resource Load Balance
Optimize the load balance for various resources (goods, shelves, etc.) through batching and scheduling of different tasks, to improve the overall task completion efficiency.
Use operational research algorithms to optimize the storage and storage area of the goods and improve the latency and throughput of our overall operation.
Dynamically configures the overall operating rhythm and various optimizations based on current robot status, item status, order status and order forecasts.