Args in file facemodel.py

[ThanhNhann]

Args in file facemodel.py

[1996scarlet]

The first thing I want to declare is that this project is based on InsightFace.
I am sorry that I have not updated readme yet.
The args parser function is defined as def start_up_init(train_mode=False) in helper.py

def start_up_init(train_mode=False):
    parser = argparse.ArgumentParser(description='Arc Face Online Test')

    # =================== General ARGS ====================
    if not train_mode:
        parser.add_argument('ip_address', type=str,
                            help='相机的IP地址或测试用视频文件名')
    parser.add_argument('--face_recognize_threshold', type=float,
                        help='可疑人员识别阈值', default=0.95)
    parser.add_argument('--max_face_number', type=int,
                        help='同时检测的最大人脸数量', default=8)
    parser.add_argument('--max_frame_rate', type=int,
                        help='最大FPS', default=25)
    parser.add_argument('--image-size', default='112,112',
                        help='输入特征提取网络的图片大小')
    parser.add_argument('--dangerous_threshold', type=int,
                        help='1/2报警窗口长度', default=16)
    parser.add_argument('--model', default='./model-r100-ii/model,0',
                        help='特征提取网络预训练模型路径')
    parser.add_argument('--gpu', default=0, type=int,
                        help='GPU设备ID，-1代表使用CPU')
    parser.add_argument('--det', default=0, type=int,
                        help='设置为1代表使用R+O网络进行检测, 0代表使用P+R+O进行检测')
    parser.add_argument('--flip', default=1, type=int,
                        help='是否在训练时进行左右翻转相加操作')
    parser.add_argument('--threshold', default=1.24, type=float,
                        help='空间向量距离阈值')
    parser.add_argument('-v', '--video_mode', action="store_true",
                        help='设置从视频读取帧数据', default=False)
    parser.add_argument('-c', '--cv_test_mode', action="store_true",
                        help='设置本地预览', default=False)
    parser.add_argument('--mtcnn_minsize', type=int,
                        help='mtcnn最小检测框的尺寸（越小检测精度越高）', default=50)
    parser.add_argument('--mtcnn_factor', type=float,
                        help='mtcnn图像缩放系数（关联图像金字塔层数，越大检测精度越高）', default=0.709)
    parser.add_argument('--mtcnn_threshold', type=float, nargs='+',
                        help='mtcnn三层阈值', default=[0.6, 0.7, 0.92])

    return parser.parse_args()

And the entrance of this system is mikomiko.py

To launch this system you'll need node-js and python3.7.

In this system we use Node-js as the websocket server to transfer image data from python client to webpage client.

I will update the project in the next few days. Thank you for your interest in this project.

[ThanhNhann]

Thank you for your explain, it's very helpful for me. In this project, will you use a search algorithm to compare an image input with the dataset?
And one more question is in the function def start_up_init(train_mode=False) in helper.py: how to unstantd the argument in each element parser.add_argument().

[1996scarlet]

In this system we have to load 3 models
1、Pre-trained MTCNN model for face detection (provided by others)
2、Pre-trained ArcFace model for face embedding (provided by others)
3、Pre-trained MLP model for face recognition (provided use train.py)

Once we have detected a face, put it into ArcFace model to get 512 dims embedding vector
Then put the vector into MLP model to get prediction result

async def embedding_loop(preload):
    # =================== FR MODEL ====================
    mlp, class_names = read_pkl_model('./model-mlp/mlp.pkl')
    preload.gpu = -1
    embedding = face_embedding.EmbeddingModel(preload)
    while True:
        img = suspicion_face_queue.get()
        dt = time.strftime('%m-%d %H:%M:%S')

        predict = mlp.predict_proba([embedding.get_one_feature(img)])
        prob = predict.max(1)[0]

        result_queue.put((img, dt, prob, class_names[predict.argmax(1)[0]]))

=========

Some argument have been deprecated, I will tell you the currently availables:

ip_address we use a web camera interface to get the frame. You can change the code to use a local video as the input.
--max_face_number the max number of face that you want MTCNN to detect
--max_frame_rate the gap time to get frame is based on this argument.(see mikomiko.py line 165)
--image-size input image size of ArcFace.
--model the path of ArcFace Model
--gpu choose the Device ID of gpu to use, -1 for use cpu
--flip whether to do flip in the MLP train process
--mtcnn_minsize ,--mtcnn_factor,--mtcnn_threshold are arguments for MTCNN, for the minisize of face to detect, the pyramids of image to factor and the threshold for P-NET R-NET O-NET

The system can be CPU real-time (upload image with max_frame_rate), but it cost nearly 3GB memory.
Use GPU can reduce the lantency to only 0.2s. (change args.fill_number to 6 at mikomiko.py line 185)

[ThanhNhann]

Thank you very much, I hope to see the next update soon 💯

[anguoyang]

@1996scarlet 用webcam试了一下，我有一张照片和stalin的相似度高达99%：）