dulimov/AuraFace-v1-rk3588

🔷 AuraFace-RKNN — Face Recognition Model for RK3588 📦 Format: .rknn 🎯 Purpose: Extract 512-dim facial embeddings 🧠 Architecture: ResNet100 + ArcFace 🚀 Target: Rockchip RK3588 NPU

📥 Source Base ONNX model: fal/AuraFace-v1

Repository: fal/AuraFace-v1

Conversion done using: rknn-toolkit2

🔧 How this was converted Downloaded glintr100.onnx from fal/AuraFace-v1

Quantized using 50 face images from mwong/faces

Converted to .rknn using rknn.build(...) with int8 quantization

⚙️ Inference Example

from rknn.api import RKNN import numpy as np import cv2

rknn = RKNN() rknn.load_rknn('auraface.rknn') rknn.init_runtime(target='rk3588')

Prepare input image (BGR, 112x112)

img = cv2.imread('face.jpg') img = cv2.resize(img, (112, 112)) img = img.transpose(2, 0, 1)[np.newaxis, :] # Shape: [1, 3, 112, 112]

Run inference

outputs = rknn.inference(inputs=[img]) embedding = outputs[0][0]

Normalize embedding (recommended for comparison)

embedding /= np.linalg.norm(embedding) 🧪 Output 512-dimensional float32 embedding vector

Ready for face similarity checks (e.g. cosine or L2 distance)

📄 License Based on the license of the upstream model (Apache 2.0 or equivalent from fal/AuraFace-v1).

📌 About this repository This RKNN version was built for Rockchip RK3588-based devices (e.g. Orange Pi 5 Pro) to enable fast and efficient on-device face recognition using NPU.