ArcFace glintr100 — Face Embedding (ONNX)
512-dimensional face-embedding model used for face recognition in the PULAO vision pipeline.
⚠️ Provenance / license — read before reuse. This matches InsightFace's ArcFace
glintr100(ResNet-100 trained on Glint360K). InsightFace's pretrained models are released for non-commercial research purposes only. This repository is private; do not redistribute or use commercially without permission from the original authors. Source: deepinsight/insightface.
Files
glintr100.onnx— 260.7 MB- SHA-256:
4ab1d6435d639628a6f3e5008dd4f929edf4c4124b1a7169e1048f9fef534cdf
Inputs / outputs
- Input
input.1: float32[1, 3, 112, 112], RGB, CHW. - Output:
[1, 512]L2-comparable face embedding. Compare identities with cosine similarity.
Usage (onnxruntime)
import cv2, onnxruntime as ort
sess = ort.InferenceSession("glintr100.onnx", providers=["CPUExecutionProvider"])
face = cv2.resize(crop, (112, 112))
face = cv2.cvtColor(face, cv2.COLOR_BGR2RGB).transpose(2, 0, 1).astype("float32") / 255.0
emb = sess.run(None, {"input.1": face[None]})[0][0] # shape (512,)
Preprocessing note: this pipeline normalizes with
/255. Canonical ArcFace preprocessing is(x - 127.5) / 128. Use whatever your enrolled gallery was built with — mixing the two degrades similarity scores.
Intended use
Non-commercial research / prototyping of face recognition only.