Blog

UniFace ships two face parsers that answer different questions. I ran both on the same faces to find where the class map earns its keep, how small a face can get before it falls apart, and why the bigger backbone is not the safer pick.

MiniFASNet in UniFace tells a live face from a picture of one. A live capture, a print, a screen replay and old photographs, judged in under a millisecond.

UniFace ships five face recognition families. I ran them all on the same old photographs to see which to pick, and why one model's threshold won't fit another.

MediaPipe Face Mesh in UniFace on ONNX Runtime: 468 dense 3D landmarks from a 192x192 crop, with BlazeFace bundled and parity checks against mp.solutions.

FaceAttribNet in UniFace predicts eye openness, eyeglasses, sunglasses, and mask attributes from a 128x128 face crop using five independent binary heads.

Practical Python interview answers on data structures, object identity, memory management, the GIL, generators, decorators, and async FastAPI behavior.

UniFace is a Python library for face analysis on ONNX Runtime: six detectors, five recognition models, 468-point mesh, parsing, attributes, and quality scoring.

Face parsing with BiSeNet, ResNet18 and ResNet34 backbones, CelebAMask-HQ training, PyTorch inference, and ONNX export.

Tiny-Face compares SlimFace, RFB, and compact RetinaFace variants for small face detection models with WIDER FACE results and PyTorch/ONNX weights.

A RetinaFace implementation with MobileNet and ResNet backbones, WIDER FACE evaluation, webcam inference, ONNX support, and pretrained weights.

Why geometric intuition for vectors and matrix transformations matters in machine learning, with practical examples and recommended learning resources.

MobileGaze estimates gaze direction with ResNet, MobileNet, and MobileOne backbones. Gaze360-trained, with PyTorch and ONNX weights for real-time inference.

Head pose estimation with ResNet and MobileNet backbones, SCRFD face detection, PyTorch and ONNX weights, and pitch/yaw/roll output for real-time video.