Projects

Open-source work, mostly computer vision and mostly faces. The pattern across all of it is the same: take a model that works in a paper or a closed SDK, reimplement it cleanly, verify it against the original, and ship it as something you can pip install and run on CPU.

UniFace

A unified face analysis library for Python

Detection, recognition, landmarks including FaceMesh (MediaPipe), parsing, matting, tracking, and attributes spanning age, emotion, gaze, head pose, and anti-spoofing. Every model returns the same Face object, so one stage feeds straight into the next.

GitHub Docs Live demo Read the write-up

Face Parsing

Real-time facial semantic segmentation with BiSeNet

Per-region face masks across 19 classes, with ResNet-18 and ResNet-34 backbones. Training code, evaluation, and ONNX export included.

GitHub Write-up

Gaze Estimation

Lightweight gaze direction with MobileOne

Pitch and yaw gaze estimation trained on Gaze360, with ResNet, MobileNet and MobileOne backbones for edge deployment.

GitHub Write-up

Head Pose Estimation

Yaw, pitch and roll from a single face crop

Head orientation with ResNet and MobileNet backbones, trained on 300W-LP and evaluated on AFLW2000. PyTorch training and ONNX Runtime inference.

GitHub Write-up

RetinaFace

Single-stage face detection with five landmarks

RetinaFace in PyTorch with MobileNet and ResNet backbones: training, WIDER FACE evaluation, ONNX export and pretrained weights.

GitHub Write-up

Face Anti-Spoofing

Live face or a picture of one, in under a millisecond

Minimal PyTorch and ONNX Runtime inference for Silent-Face-Anti-Spoofing (MiniFASNet). Catches prints, screen replays and paper masks from a single face box.

GitHub Write-up

Face Recognition

MobileFace and SphereFace, PyTorch to ONNX Runtime

Recognition models trained on MS1MV2 with MobileNet and Sphere backbones, exported to ONNX. Two of the five families UniFace ships.

GitHub Write-up

More on GitHub, including smaller tools, tutorials, and dataset utilities.