CWT LAB / ATTRIBUTION
Built on open-source work.
Human by Vladimir Mandic
This exhibit uses the Human library and live demo, created by Vladimir Mandic. Human provides the face, body, hand, expression, and gesture inference and drawing capabilities. Copyright © Vladimir Mandic. Used under the MIT License.
The demo JavaScript and helper files are adapted from upstream commit d0c4c83155fd2325dbb4725ffcaf0ed0109f319c. The runtime and default models are loaded from the pinned npm package @vladmandic/human@3.3.6 through jsDelivr. The optional single and group sample images are from the same pinned Human repository revision.
Models and underlying software
Human builds on TensorFlow.js and models from multiple authors, including Google MediaPipe and MoveNet and the Oarriaga emotion model. See Human’s model documentation and attribution for each model’s source and applicable licensing. The Human software license does not replace individual model licenses.
Multi-person body pose uses Google MoveNet MultiPose from Human Models revision bc66dc53bac03c96d35a7e6daaf717e72f3985f5, loaded through jsDelivr. This model supports up to six body poses.
CWT Lab exhibit adaptation
Presented by the Construction Workforce & Technology Lab. Brand colors and wordmark treatment follow the lab’s website. The branded shell, exhibit instructions, fullscreen mode, stop/reset control, and explicit camera launch were added for this exhibit. The original demo’s model and display controls remain available.
Adaptations also include pinned asset locations, responsive styling, face/body/hand defaults, disabled service-worker registration, and camera-start and error-reporting fixes. Face-description analysis (which includes age and gender estimates) is off by default.
The subject limit defaults to four and can be set from one to six. Face and body capacity follow that limit; hand capacity is twice the limit. Changing the limit restarts the demo. Numbered face, pose, and hand detections are independent and can change between frames; they do not establish identity or reliably associate a hand with a particular face. Visibility, overlap, distance, lighting, and device performance affect results.
Processing and interpretation
Camera frames and selected images are analyzed in the visitor’s browser. This adaptation does not record or upload them. Loading the site, library, models, and sample image makes ordinary requests to GitHub Pages and jsDelivr. Internet access is required to load those resources.
Expression categories are model estimates, not measurements of internal emotional state. This is an exploratory demonstration.