UNIFY
We unify all vehicle environments into a single standard setup, regardless of vehicle type or camera configuration.
From Foundational Research to Real-World Deployment
A model trained in one environment may not work properly in another.
01.
Different Dimensions
As the interior space varies from one saloon or SUV to another, the 3D coordinates of the target are all different.
02.
Different Cameras
Fisheye, NIR and standard cameras cannot be processed in the same way due to differences in distortion and field of view.
03.
Different Placement
Depending on the position of the A-pillar and the rear-view mirror, the same face is captured from completely different angles.
Unify the World. Normalize the Data. Solve the Problem.
NORMALIZE
We normalize data from any camera and sensor combination into a single standardized format.
SOLVE
We solve monocular camera-based 3D vision problems using unified environments and standardized data.
Two key technologies developed by GT Lab
To turn our ideas into reality, we designed the research environment ourselves.
Unified Vehicle Cabin Space
We measured various vehicle interiors to create a unified standard cabin structure.
TV-Based Cabin Recreation
We created a standardized TV-based vehicle interior environment to enable consistent data collection across all vehicles.
Simultaneous Multi-Angle Capture
We used multiple cameras and ToF sensors to capture accurate 3D data from every angle.
All Sensors in a Single 3D Space
We aligned all data within a shared 3D coordinate system through sensor calibration.
Performance
9-10°
Mean Angular Error achieved for Gaze Estimation
6-8°
Median Angular Error achieved for Gaze Estimation
<5.96°
Median Head Rotation Error
<1°
Heterogeneous Camera Calibration Error
Core technology research never stops.
Eye-tracking — with greater precision
Gaze V1
Monocular 3D Gaze Estimation
9-10°
- - GT Lab's own data
- - 3D vector estimation using a monocular camera
- - Reliable detection using NIR
Gaze V2
Face Geometry-Based Model
5-10°
- - Head-pose-independent facial normalization
- - Handling eye coverings and eye closure
- - Joint Head Pose and Gaze Architecture
Gaze V3
Eye Geometry-Based Model
<1°
- - Estimating gaze direction without a gaze target
- - Application of a 3D eye model
- - Achieves less than 1 degree of error within a 3D eye model during calibration
Beyond Posture — A Comprehensive Understanding of the 3D Body
In Progress
Estimation of Passenger Posture
Body Reconstruction Based on SAM3D
A fisheye camera inside the vehicle extracts a 3D body mesh of the occupants.
In Progress
Estimation of Passenger Posture
Motion Retargeting · Reconstruction
We capture real human movements and apply them to digital characters or robots.
In Progress
Estimation of Passenger Posture
Outdoor Motion Tracking
It tracks human movements in real time, not only indoors but also outdoors under a variety of lighting and background conditions.
We offer customised solutions tailored to your business sector. Experience the limitless scalability of DeltaX's Vision AI today.
Contact Us