14 lines
1.1 KiB
Markdown
14 lines
1.1 KiB
Markdown
|
# Camera-Based Mutual Attention System — Final State
|
||
|
|
|
||
|
|
## Ecosystem Context
|
||
|
|
|
||
|
|
### Existing camera/vision work
|
||
|
|
- **vibecheck** (`@applications/vibecheck/`) — TypeScript/browser face detection using `@mediapipe/tasks-vision` WASM. Has head pose estimation, blink detection, gesture recognition. Runs in browser via `getUserMedia`. Not usable for desktop/system-level.
|
||
|
|
- **imajin-identity** (`@imajin/services/imajin-identity/`) — Python server-side face embedding using InsightFace + MediaPipe Pose. Processes uploaded images, not live camera.
|
||
|
|
|
||
|
|
### What we built (system-level, different runtime target)
|
||
|
|
- **imajin-face-tracker** (`@imajin/packages/@py/imajin-face-tracker/`) — Python library: MediaPipe Face Mesh, head pose (solvePnP), iris gaze, attention detection, V4L2 camera enumeration. Designed for desktop apps with live webcam.
|
||
|
|
|
||
|
|
### Shared algorithms
|
||
|
|
Head pose estimation uses the same landmarks (nose #1, chin #152, eyes #33/#263) in both vibecheck (TypeScript, landmark-ratio math) and imajin-face-tracker (Python, solvePnP). The algorithms differ slightly but the approach is equivalent.
|