Data collection & annotation
Our expert team collects diverse datasets and uses in-house tools for complex annotations, including detailed landmarks, segmentation masks, and attribute labels.
DEEPIXEL
Full Stack
Vision AI Company
We are a pioneering computer vision startup, redefining how machines see and understand the world.
Human-XOn-device intelligence
One camera. A deeper understanding.
01 / About us
From Data to
On-Device Intelligence
Deepixel develops computer vision AI across the full pipeline, from data collection, annotation, and augmentation to deep learning model architecture, training, and optimization.
Our team of experts uses in-house annotation tools to handle complex labeling tasks across diverse datasets. Combined with data augmentation, this work supports models for human detection, tracking, segmentation, and analysis under real-world conditions.
Our Human-X engine brings this work onto devices. We continually improve inference speed, memory use, and computational efficiency to deliver accurate, real-time AI on resource-constrained hardware.
Data augmentation
Expand training variation in lighting, pose, and appearance to improve model robustness and generalization.
Deep learning architecture
Our FEWER-Net architecture enriches essential features to balance recognition accuracy with efficient computation. Explore FEWER-Net.
Model optimization
Refine training and inference to reduce model size and computational cost while preserving accuracy.
On-device performance
Accelerate CPU inference with XNNPACK and SIMD on supported platforms, while optimizing memory use and keeping data on the device.
Our deep learning architecture
FEWER-Net.
Fewer features. Richer information.
Feature Enrichment With Efficient Refinement
Deepixel’s FEWER-Net enriches the core features essential to each task, increasing the useful information carried by each feature to make effective use of a compact feature set.
Lightweight convolutions generate complementary features, while channel attention emphasizes relevant information to learn richer representations.
By projecting enriched features back into fewer channels, the repeated cycle of expansion and refinement enables small model sizes and low memory demands—with models as small as hundreds of kilobytes and inference times of only a few milliseconds on mobile CPUs.
-
01 / Input
Visual features
Core visual information for the task.
-
02 / Enrich
Lightweight convolution
Generate complementary features efficiently.
-
03 / Focus
Channel attention
Emphasize the information that matters.
-
04 / Refine
Projection / refinement
Project enriched features into fewer channels.
-
05 / Output
Compact & enriched output
Richer representations with a smaller footprint.
Repeat expansion and refinement to retain useful information in fewer channels.
Efficiency, by design
Small footprint. Strong performance.
FEWER-Net vs. MobileNet v5
- Parameters
- 96%fewer parameters
- Inference time
- 50%less inference time
- Speed
- 180%faster
- Performance
- +15%performance gain
02 / Human-X
On-device Real-time AI Engine for Full Human Understanding
Detection, Tracking, Segmentation, Measurement and Analysis
Human-X turns camera images into landmarks, poses, masks, and analysis results. Its task-specific modules provide the visual information needed for interactive applications, from virtual try-on to human-machine interfaces.
Platform agnostic
Deploy across PCs, Android, iOS, web browsers, and edge devices, with AI processing kept on-device.
CPU-based AI
Runs AI inference without a dedicated GPU; 3D rendering is handled separately.
Efficient models
Task-specific architectures balance accuracy, memory use, and inference speed.
Focused capabilities
Dedicated modules provide the outputs each application needs.
01 / Experience
AI for new digital experiences
AR Virtual Try-ons, Beauty AI, Skin Analysis, Image Filters
02 / Movement
AI for Human Behavior
Health Monitoring, Action Recognition, Sports Tech., Smart Fitness
03 / Interaction
AI for intelligent interface
Human Metadata Recognition, Physical AI, Human–Robot Interaction (HRI), Human–Computer Interaction (HCI)
See it in motion
Vision that keeps up with real life.
Watch Human-X recognize and follow movement in real time from a single camera feed.
Algorithm modules
Computer vision algorithms.
Explore all modules for detection, tracking, segmentation, measurement, and analysis.
01+NEXT_MODULE
Face detection
Face bounding boxes and detection confidence.
Open module ↗ 02Face landmarks
108-point, 282-point, or mesh output with head pose.
Open module ↗ 03Face attributes
Face shape and estimated age from validated facial landmarks.
Open module ↗ 04Pupillary distance
Distance estimate in millimeters with measurement status.
Open module ↗ 05Ring-finger tracking
Dedicated ring-finger points, mask, and cylinder pose.
Open module ↗ 06Hand landmarks
21-point skeletal tracking in real time.
Open module ↗ 07Wrist tracking
Wrist mask, cylinder pose, and tracking confidence.
Open module ↗ 08Upper-body segmentation
Upper-body masks for foreground compositing.
Open module ↗ 09Face & skin segmentation
Facial skin masks for cosmetics and surface effects.
Open module ↗ 10Face skin analysis
Capture a frame to explore skin scores and six analysis overlays.
Open module ↗ 11Hair segmentation
Hair-region masks for virtual color effects.
Open module ↗ 12Foot detection
Left and right foot masks with 3D poses.
Open module ↗ 13Full-body landmarks
33-point body tracking with estimated 3D world landmarks.
Open module ↗More in motion
We are continually expanding what Human-X can understand.
Use cases
Try it with your camera.
These examples show how our core computer vision algorithms power real-world applications. StyleAR, one of our product lines, brings this technology to augmented reality virtual try-on.
01
Eyewear
Facial landmarks and 3D eyewear placement.
Try on ↗ 02Ring
Ring-finger pose and occlusion.
Try on ↗ 03Watch
Wrist pose and watch occlusion.
Try on ↗ 04Footwear
Left and right foot pose with occlusion.
Try on ↗ 05Makeup
Facial tracking and layered makeup effects.
Try on ↗ 06Hair
Hair segmentation and live color effects.
Try on ↗ 07Earrings
Facial tracking and paired earring placement.
Try on ↗ 08Skin tone
Target-color tint with adjustable opacity on detected skin.
Try on ↗Product sets
Style them together.
Try a coordinated pair in one camera view. Find your fit with eyewear and earrings, or complete your hand styling with a bracelet and rings.
Understanding Human-X
About the technology.
What is Human-X?
Human-X is Deepixel’s on-device computer vision engine. It turns camera images into information applications can use: the locations of facial and hand landmarks, the movement of body regions, segmentation masks, and visual attributes. Each module is designed for a specific task, with deep learning architectures optimized for accuracy and efficient inference. Explore the Human-X modules to see the capabilities in action.
Does Human-X require a GPU, and can it run in a browser?
Human-X runs AI inference on the CPU without requiring a dedicated GPU, and its modules can run directly in a compatible browser. Lightweight models, XNNPACK-optimized operators, and SIMD vector processing help accelerate inference on supported platforms. Actual speed depends on the hardware, selected module, and input resolution. For applications such as virtual try-on, 3D graphics are rendered separately using WebGL and may use the device’s GPU.
How are Deepixel, Human-X, and StyleAR related?
Deepixel develops the technology, from data collection, annotation, and augmentation to model architecture and on-device optimization. Human-X is the core computer vision engine that makes these capabilities available to applications. StyleAR is one of our product lines, applying this engine to augmented reality virtual try-on for eyewear, jewelry, watches, footwear, and beauty products. Explore the application demos or visit StyleAR to learn more.
How can I try the algorithms or discuss an integration?
Explore the individual modules to evaluate specific computer vision capabilities, or try the application demos to see them in a complete experience. Camera demos require a compatible browser, a secure connection, and camera permission. For an integration discussion, share your use case, target devices, and performance requirements with deepixel@deepixel.xyz. We can discuss suitable modules, deployment options, and licensing.
03 / For developers
Explore the technology on GitHub.
Read Deepixel’s public integration examples, algorithm documentation, and evaluation tools. These are related developer resources; APIs and model outputs may differ from the browser demos. Check each repository for licensing and platform requirements.
Face landmark tracking ↗
Facial keypoints, head pose, and Python integration examples.
Wrist landmark tracking ↗
Wrist landmarks, orientation, and wearable alignment concepts.
Upper-body segmentation ↗
Portrait masks and foreground separation documentation.
Segmentation evaluation ↗
Tools for measuring mask overlap and boundary quality.
TensorFlow Lite benchmarking ↗
A guide to measuring model inference performance.
StyleAR integration guide ↗
PlayCanvas setup and virtual try-on implementation examples.
04 / Contact us