SynchroCV
SynchroCV is the production implementation of published research, built after co-authoring "Transfer Learning-Based Distance-Adaptive Global Soft Biometrics Prediction in Surveillance" in Electronics (MDPI), 2025. It runs a 6-stage ML pipeline live in a browser: face detection, distance estimation, image quality scoring, adaptive contrast enhancement, gender prediction, and distance-penalised confidence adjustment. The system is honest about its own uncertainty: a far, blurry face legitimately gets a low confidence score rather than a false 90%.
Results & Impact
The Challenge
Face recognition collapses at distance in browser environments because browsers deliberately hide hardware camera parameters (focal length, sensor size) to prevent device fingerprinting, making standard physics-based distance formulas impossible. Additionally, off-the-shelf models trained on studio portrait datasets perform poorly on real webcam footage at 7–10 metres.
The Solution
Built a calibration-free face-ratio heuristic for distance estimation, trained a custom 3-model ensemble (ResNet50+Attention, MobileNetV2, Multi-Scale) on 202K CelebA images with synthetic distance degradation (downscale→upscale→blur→noise per distance category) to bridge the studio-to-webcam domain gap. Used ArcFace Loss to push class clusters apart so noisy far-face features are less likely to cross the decision boundary. Streamed frames over binary WebSocket (not Base64) saving 33% bandwidth. Discovered and fixed an undocumented InsightFace bug: passing tight face crops silently returns wrong high-confidence predictions because the model needs full-image context (hair, shoulders); fix was passing the full frame.