For each face image, the model produces a fixed length embedding vector corresponding to the face in the image. The vectors from face images of a single person have a higher similarity than that from different persons. Therefore, the model is primarily used for face recognition/verification. It can also be used in other applications like facial feature based clustering.
ArcFace is a CNN based model for face recognition which learns discriminative features of faces and produces embeddings for input face images. To enhance the discriminative power of softmax loss, a novel supervisor signal called additive angular margin (ArcFace) is used here as an additive term in the softmax loss. ArcFace can use a variety of CNN networks as its backend, each having different accuracy and performance.
The model LResNet100E-IR is an ArcFace model that uses ResNet100 as a backend with modified input and output layers.
| Model | Download | Download (with sample test data) | ONNX version | Opset version | LFW * accuracy (%) | CFP-FF * accuracy (%) | CFP-FP * accuracy (%) | AgeDB-30 * accuracy (%) |
|---|---|---|---|---|---|---|---|---|
| LResNet100E-IR | 248.9 MB | 226.6 MB | 1.3 | 8 | 99.77 | 99.83 | 94.21 | 97.87 |
| LResNet100E-IR-int8 | 63 MB | 46 MB | 1.13.1 | 11 | 99.80 |
Compared with the fp32 LResNet100E-IR, int8 LResNet100E-IR accuracy drop ratio is 0% and performance improvement is 1.78x in LFW dataset.
The performance depends on the test hardware. Performance data here is collected with Intel® Xeon® Platinum 8280 Processor, 1s 4c per instance, CentOS Linux 8.3, data batch size is 1.
* each of the accuracy metrics correspond to accuracies on different validation sets each with their own validation methods.
We used MXNet as framework to perform inference. View the notebook arcface_inference to understand how to use above models for doing inference. A brief description of the inference process is provided below:
The input to the model should preferably be images containing a single face in each image. There are no constraints on the size of the image. The example displayed in the inference notebook was done using jpeg images.
In order to input only face pixels into the network, all input images are passed through a pretrained face detection and alignment model, MTCNN detector. The output of this model are landmark points and a bounding box corresponding to the face in the image. Using this output, the image is processed using affine transforms to generate the aligned face images which are input to the network. Check face_preprocess.py and inference notebook for code.
The model outputs an embedding vector for the input face images. The size of the vector is tunable (512 for LResNet100E-IR).
The post-processing involves normalizing the output embedding vectors to have unit length. Check face_postprocess.py for code.
To do quick inference with the model, check out Model Server.
Refined MS-Celeb-1M is a refined version of the MS-Celeb-1M dataset. The refined version contains 3.8 million images from 85000 unique identities.
The following three datasets are used for validation:
-
Labelled Faces in the Wild (LFW) contains 13233 web-collected images from 5749 identities with large variations in pose, exposure and illuminations.
-
Celebrities in Frontal Profile (CFP) consists of 500 subjects, each with 10 frontal and 4 profile images.
-
Age Database (AgeDB) is a dataset with large variations in pose, expression, illuminations, and age. AgeDB contains 12240 images of 440 distinct subjects, such as actors, actresses, writers, scientists, and politicians. Each image is annotated with respect to the identity, age and gender attribute. The minimum and maximum ages are 3 and 101, respectively. The average age range for each subject is 49 years. There are four groups of test data with different year gaps (5, 10, 20 and 30 years respectively). The last subset, AgeDB-30 is used here as its the most challenging.
- Download the file
faces_ms1m_112x112.zip: 8.1 GB - Unzip
faces_ms1m_112x112.zipto produce a folder of the same name. Use path to this folder in the notebooks. This folder contains the training as well as the validation datasets.
The accuracies obtained by the models on the validation set are mentioned above. Maximum deviation of 0.2%(CFP-FP) in accuracy is observed compared to that in the paper.
We used MXNet as framework to perform training. View the training notebook to understand details for parameters and network for each of the above variants of ArcFace.
The validation techniques for the three validation sets are described below:
-
LFW : Face verification accuracy on 6000 face pairs.
-
CFP : Two types of face verification - Frontal-Frontal (FF) and Frontal-Profile (FP), each having 10 folders with 350 same-person pairs and 350 different person pairs.
-
AgeDB : Validation is only performed on AgeDB-30 as mentioned above, with a metric same as LFW.
We used MXNet as framework to perform validation. Use the notebook arcface_validation to verify the accuracy of the model on the validation set. Make sure to specify the appropriate model name in the notebook.
LResNet100E-IR-int8 is obtained by quantizing fp32 LResNet100E-IR model. We use Intel® Neural Compressor with onnxruntime backend to perform quantization. View the instructions to understand how to use Intel® Neural Compressor for quantization.
Download model from ONNX Model Zoo.
wget https://github.com/onnx/models/raw/main/vision/body_analysis/arcface/model/arcfaceresnet100-8.onnxConvert opset version to 11 for more quantization capability.
import onnx
from onnx import version_converter
model = onnx.load('arcfaceresnet100-8.onnx')
model = version_converter.convert_version(model, 11)
onnx.save_model(model, 'arcfaceresnet100-11.onnx')cd neural-compressor/examples/onnxrt/body_analysis/onnx_model_zoo/arcface/quantization/ptq_static
bash run_tuning.sh --input_model=path/to/model \ # model path as *.onnx
--dataset_location=/path/to/faces_ms1m_112x112/task.bin \
--output_model=path/to/save- All models are from the paper ArcFace: Additive Angular Margin Loss for Deep Face Recognition.
- Original training dataset from the paper MS-Celeb-1M: A Dataset and Benchmark for Large-Scale Face Recognition.
- InsightFace repo, MXNet
- Intel® Neural Compressor
- abhinavs95 (Amazon AI)
- mengniwang95 (Intel)
- yuwenzho (Intel)
- airMeng (Intel)
- ftian1 (Intel)
- hshen14 (Intel)
Apache 2.0