Model FAQs
How is Theia's model trained?
Theia3D has been trained on millions of images from publicly available and proprietary datasets.
These images include a massive diversity of humans, with variation along any imaginable characteristic such as age, weight, height, skin tone, clothing type, clothing colour, and others. The images are also diverse with respect to the setting, background, pose of the person(s), image source, and other factors.
Images are annotated by an expert annotator, with Quality Assurance steps including review by at least one other expert annotator.
What landmarks are used?
Theia3D tracks 124 keypoints across the entire body (Theia3D Axiom v2024 and onwards), providing sufficient points for all segments to be fully observable in 3D space.
We are unable to provide the specific keypoints that are tracked; however, they include, but are not limited to, common anatomical landmarks that are typically used in biomechanics and motion capture contexts as well as other technical keypoints to increase redundancy and improve tracking in the face of occlusion or other challenges.
How are the model segments defined?
Theia3D uses 17 segments to track the movement of the body, as described in the Default Model Description page.
These segments are defined to have biomechanical and anatomical relevance, and will provide sufficient data for most users' applications.
Can I build custom segments?
The 6-DOF pose of all segments are exported as 4x4 pose matrices within the selected output file format (e.g. C3D, FBX, JSON), along with a select few landmarks for relevant segments (e.g. left and right heels).
When using post-processing software such as Visual3D, it is possible to build custom segments by defining virtual landmarks relative to the existing segment pose matrices. However, since the raw tracked landmarks are not included as output signals, these custom segments are typically limited to having a different static orientation but tracking the identical movement as the existing output segments using the 4x4 pose matrices.
What is the difference between pelvis_4X4 and pelvis_shifted_4X4 signals?
For all versions of Theia3D, these pose matrix signals express the position and orientation of the pelvis segment. However, there are some very slight differences that may be present, depending on the version of Theia3D used:
For Theia3D v2023 and earlier, these two sets of signals are nearly identical, with the only difference being that the pelvis_shifted_4x4 signal has its origin located at the height of the pelvis plane, whereas the pelvis_4x4 signal has its origin located at the midpoint of the hip joint centers. Otherwise, these signals are identical, and if you were to calculate a vector between their origins you would find it to have a constant length and constant direction within the pelvis reference frame. These separate signals were used to define the proximal (pelvis_shifted_4x4) and distal (pelvis_4x4) ends of the pelvis segment in order to construct the pelvis segment in Visual3D and other similar post-processing software.
For Theia3D Apollo v2024 and later, these signals are identical and are a holdover from previous versions. The pelvis_4x4 origin is now located at the pelvis plane (i.e. where the pelvis_shifted_4x4 signal was previously), and is used to define the proximal end of the pelvis, while the distal end is defined as the midpoint between the hip joint centers which are coincident with the thigh segments origins. Assuming an up-to-date version of Visual3D is used, the pelvis segment will be constructed accordingly.
What is the abdomen segment?
The abdomen segment in the full-body model does not have its own tracked points, it is scaled and positioned based on the pelvis and thorax through inverse kinematics. The abdomen segment primarily mainly exists to bridge those two segments and provide a continuous chain for FBX/animation export.
Functionally:
It allows some independent x-y tilt (flexion/extension, ab/adduction) relative to the pelvis.
It does not capture z-axis rotation (internal/external rotation).
Its pose should not be interpreted as a true lower-back segment; it’s more of a connector than a biomechanically distinct region.
Last updated
Was this helpful?