Deepixel is a visual-intelligence deep-tech startup specializing in real-time human understanding algorithms using a single RGB camera.

We design and optimize proprietary computer vision and machine learning pipelines that perform robust 3D pose estimation, high-precision landmark detection, and temporal tracking across mobile and web environments.

Deepixel operates a fully in-house vision R&D stack, covering raw data acquisition, custom annotation pipelines, deep learning model design and training, and platform-specific optimization. This end-to-end ownership allows us to tightly couple data, models, and inference pipelines, ensuring algorithmic efficiency, stability, and accuracy under real-world constraints, and enabling solutions that are precisely tailored to the needs of each vision problem and deployment platform.

Deepixel’s vision stack is built around the following principles:

• Single-Camera Geometry
  Inferring 3D structure and pose from monocular RGB input

• Real-Time Inference
  Low-latency pipelines optimized for mobile CPUs, GPUs, and NPUs

• Robustness in the Wild
  Designed to handle occlusion, motion blur, illumination changes, and extreme poses

• Lightweight Models
  Architectures optimized for on-device inference without cloud dependency

• Dense facial landmark detection
• 3D head pose estimation (rotation & translation)
• Ear landmark inference under partial occlusion
• Stable temporal tracking for AR alignment

• 21-keypoint hand skeleton estimation
• Per-joint confidence and visibility modeling
• Optimized for fast motion and self-occlusion
• Suitable for gesture recognition and fine-grained interaction

• Wrist-specific landmark topology
• 3D wrist orientation estimation
• Stable tracking under rotation and partial visibility
• Designed for watch and bracelet alignment

• Multi-joint human pose understanding
• Robust keypoint localization under self-occlusion
• Temporal smoothing for stable motion tracking
• Applicable to fitness, fashion, and HCI

• Foot landmark detection and orientation estimation
• Supports stabilization and occlusion handling
• Designed for real-time footwear AR and biomechanics use cases

• Monocular 3D reconstruction
• Learned geometric priors
• Temporal filtering & motion consistency
• Landmark-centric representations
• Low-latency & light-weight model
• ROI-based inference pipelines
• Cross-platform compatibility

These characteristics allow our models to remain accurate, fast, and stable even on constrained devices.

👉 See individual repositories for implementation details and benchmarks.

Deepixel’s technology has been recognized internationally:

• CES Innovation Awards Honoree
• KES Innovation Award
• Intelligence Start-up Award
• Backed by NAVER D2 Startup Factory (D2SF)

These recognitions reflect our strength in core vision algorithms, not just applications.

Our algorithms are designed to support:

• Augmented Reality alignment
• Spatial computing interfaces
• Human-computer interaction (HCI)
• Virtual try-on systems
• Motion analysis & tracking
• On-device AI vision systems

We welcome collaboration with:

• Computer vision / Deep learning researchers
• AR / XR engineers
• Data scientist
• Hardware & camera platform teams
• Developers building real-time vision systems

📬 deepixel@deepixel.xyz
🌐 https://www.deepixel.xyz
📍 Seoul, Republic of Korea