Generate millions of photorealistic, functional, meticulously designed synthetic individuals for robust computer vision training, ensuring privacy and overcoming data scarcity.
Our Synthetic Data Cloud easily simulates complex environments, e.g. airports, construction sites, manufacturing areas. The Omnihuman module feature enables precise control over human pose, motion, personal protective equipment, occlusions, gaze estimation, and gesture recognition data for worker tracking, hazard detection, and behavior analysis without exposing sensitive real-world data.
Omnihuman enables generation of highly realistic synthetic in-cabin monitoring data, simulating diverse occupants, behaviors, and scenarios across different vehicle types. This allows training of robust computer vision models for driver fatigue detection, passenger behavior analysis, and multi-camera systems—while ensuring privacy and minimizing the need for real-world data collection. Compliant with NCAP provisions.
Images simulated with Omnihuman deliver unique realism for security & defence AI training. The Platform generates custom human detection, crowd monitoring, and threat identification scenarios in diverse environments. Users can modify terrain, lighting, and behavior patterns to create comprehensive datasets for surveillance, rescue operations, and tracking systems while maintaining complete privacy protection.
Omnihuman enables creation of functional, realistic datasets for human detection in manufacturing environments, supporting applications like safety monitoring, worker behavior analysis and access control. With advanced simulation features and control over sensor parameters, our Platform accelerates AI model development while reducing reliance on real-world data.
Accelerate smart home AI with perfect training data tailored to any environment. Omnihuman simulates realistic humans in diverse surroundings, enabling unlimited, privacy-safe computer vision datasets. Train IoT devices instantly for any lighting, layout, or user scenario—boosting performance, accuracy, and adaptability across real-world conditions.
The Omnihuman feature of our Synthetic Data Cloud enables comprehensive control over head poses (roll, pitch, yaw), emotions (Facial Action Coding System, FACS), eye movements, and facial expressions. Moreover, the synthetic characters have fully randomizable skin texture with pores, imperfections, and makeup, allowing for successful domain adaptation.
Full body, limbs, and trunk modeling enable realistic gestures, actions, and poses, including walking, sitting, slouching, head rotation, eye opening/closing, and even falls. Additionally, our synthetic humans can interact with accessory objects (96 interactions) and display over 50 face expressions.
Each scene designed and generated using Omnihuman supports a selection of modalities, from Vis to NIR to Radar to Lidar. Additionally, our Platform utilizes proprietary algorithms for light interactions with matter resulting in realistic simulation of light interactions with skin (subsurface scattering, multilayer rendering, light reflections from surfaces).
With Omnihuman, you can generate millions of synthetic individuals on our Synthetic Data Cloud, by mixing and matching 216 unique adult identities, 6 ethnicities, 5 eye colors, 5 age ranges, and 60 variations of hair. You can also randomize appearance aspects like aging, BMI, and height. Add to that accessories, such as headwear or jewellery and you get a true hornof plenty of identities.
Omnihuman allows for adjusting human behavior to your ML training needs. You can choose from 30 emotions based on FACS (e.g., neutral, tired, goofy), interactions with 96 unique accessory elements across 24 sets. You can further fine-tune your selected behaviors with 52 action units for adults. All of that to generate humans that are as realistic and information rich as possible.
Synthetic human activities offered in Omnihuman are deeply rooted in the domain knowledge. This approach ensures that every scene and every render accurately represent real-life scenarios. For example, we know that drivers often smoke or use mobile phones when they drive and that is why we included them in Omnihuman. However, there are many more to choose from.
Our Platform and seamlessly integrates with data science and MLOps tools, and is offered as a Software-as-a-service (SaaS) access. It supports procedural image and animation synthesis with full control over dataset design and generation parameters (scene parameters, size, resolution, frame order, etc.).
Omnihuman's 3D assets were custom built from scratch, utilizing motion capture technology. Our in-house CG team meticulously adjusted each model for facial landmarks, skeleton, and range of motion, in order to maximize the teaching output of synthetic data generated with those 3D assets.
Ground truth includes a vast array pixel-perfect metadata. They include: bounding boxes, 2D/3D keypoints, semantic masks for whole body, body parts, and face elements, depth maps, heat maps, gaze vectors. Also: gender, age, BMI, skin tone/ethnicity, along with sensor parameters and camera distortions.
Director Core ML, Google. Ex-Nvidia
My team and I take a great pleasure in supporting SKY ENGINE AI, our collaboration is a win-win for both Nvidia and SKY ENGINE AI. SKY ENGINE showing what is possible with Nvidia tech and SKY ENGINE AI is leading the way with synthetic data and ML platform, which ultimately will dominate how the companies train DL models. I look forward to more collaboration opportunities.
We provide a wide array of detailed labels required for training computer vision models. They include: - Classification (Whole-image content labels) - Image-aligned 2D and 3D bounding boxes (horizontal and vertical box edges) - Object-aligned 2D and 3D bounding boxes (bounding box fits snugly about the object) - Segmentation (Pixel-level shadings of training object shapes) - Instance segmentation - 3D Keypoints - Tracks of training object movement trajectories across sequences of generated imagery
We ensure the variety human appearances and behaviors required for excellent CV algorithm training outcomes by providing a large number of modifiable character properties. They include: - 216 unique adult identities - 6 ethnicities - 5 eye colors - 5 age ranges - 60 variations of hair - 30 emotions based on FACS (e.g. Neutral, tired, goofy) - Interactions with 96 unique accessory elements in 24 sets - 52 action units describing face expression for adults
You can choose from a variety of modalities, among others: - Multispectral imagery (including RGB images) - Panchromatic imagery - Near-infrared imagery - Hyperspectral imagery - Lidar data - Motion data – electro-optical - Motion data – infrared - Ultra-wideband (UWB)
Yes, all our renders are based on physical models of light interactions with surfaces and sensors, such as microfacet models for refraction through rough surfaces [1] or Fresnel term approximations for metals [2].
No, we do not. Our 3D modeling pipelines ensure that all assets remain free from penetration, obstruction, and overlapping issues, guaranteeing high-quality, realistic, and usable 3D models.
Our images are annotated automatically on the Platform. Having complete control over the scene means you possess all the information regarding the 3D dependencies present. By eliminating manual labeling, you remove biases and inconsistencies associated with labeling.
Virtually any animation can be modelled and simulated in our Synthetic Data Cloud. It depends only on the resources and time you have to spare. Our example animations include: - Dynamic shadowing - Dynamic illumination - Vehicle activity and scenarios - Human activity and scenarios - Weather - Sensor platform movement (platform motion, fly-through behavior) - Sensor movement (camera jitter, sway, or tilting)
Aging is simulated procedurally, for increased realism and randomization.
