ORCIDGitHubLinkedIn
RSS Curriculum Vitae TR

ArcFace

Turkish equivalent: ArcFaceDomain: Computer Vision

A margin-based face-recognition objective that learns angularly separated identity embeddings on a normalized hypersphere.

Angular-Margin Objective

ArcFace normalizes feature and class-weight vectors and introduces an additive angular margin for the target identity. The objective is to make same-identity embeddings compact while increasing angular separation between identities.

The training loss is not the same thing as the runtime similarity rule. A deployed system normally produces a face embedding and compares it with cosine similarity or another calibrated score.

Deployment Boundary

Detection and alignment errors propagate into the embedding. Thresholds are also domain dependent: camera, pose, illumination, age variation and the required false-match rate matter.

My implementation-oriented comparison of export/runtime behavior is in ONNX and CTranslate2 for ArcFace Models.

Direct source: The primary paper or official specification for ArcFace is linked here for verification.

Updated:

Related concepts

Computer Vision

Related technical publications

Publications whose title or summary directly references this concept.

ONNX and CTranslate2 for ArcFace Models

An ArcFace-trained ONNX face-embedding model cannot be moved directly to CTranslate2 merely because its graph is available in ONNX; the supported model architecture and inference graph must also match.

IBM POWER9 AC922 for Digital Forensics and Artificial Intelligence

My IBM POWER9 AC922 has been a long-running engineering platform since 2019 for digital forensics and AI work involving AltiVec/VSX, OpenMP, CUDA, Tesla V100, dlib, MXNet, FAISS, ArcFace, Kaldi, Vosk, whisper.cpp, OCR, file carving, and current CTranslate2/faster-whisper experiments.