HenryOnilude/distance-recognition-live-demo
Distance-adaptive face recognition demo showcasing accuracy degradation with distance. Live webcam processing with real-time distance estimation and demographic predictions.
What's novel
Distance-adaptive face recognition demo showcasing accuracy degradation with distance. Live webcam processing with real-time distance estimation and demographic predictions.
Code Analysis
10 files read · 4 roundsA web-based face analysis demo that uses InsightFace for detection/age/gender prediction, estimates distance via a simple linear heuristic based on face-width-to-image-width ratio, and applies distance-conditioned CLAHE preprocessing and confidence scaling.
Strengths
Solid engineering on the web infrastructure side: WebSocket streaming with backpressure, exponential backoff reconnection, lazy model loading, and multi-level fallback (SCRFD → DNN → Haar). The BRISQUE-inspired quality assessment and distance-adaptive CLAHE are reasonable, if not novel, image processing techniques.
Weaknesses
The core 'distance estimation' is a trivial linear interpolation (face_ratio → 0.5-10m), not the physics-based pinhole model the README and docstrings claim. The ML training pipeline uses random noise as placeholder images instead of actual CelebA data, making the 'distance-aware model' non-functional. The README's 'Empirical Validation Results' table (89.1%, 82.3%, 72.3%) is presented as measured performance but appears to be aspirational targets hardcoded into confidence multipliers.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
78
Contributors
1
Files
55
Active weeks
15
Repository
Language
Python
Stars
2
Forks
0
License
MIT