RetinaFace.py

import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'

#---------------------------

import numpy as np
import tensorflow as tf
import cv2

import retinaface_model
import preprocess 
import postprocess

#---------------------------

import tensorflow as tf
tf_version = int(tf.__version__.split(".")[0])

if tf_version == 2:
    import logging
    tf.get_logger().setLevel(logging.ERROR)

#---------------------------

def build_model():
    
    global model #singleton design pattern
    
    if not "model" in globals():
        
        model = tf.function(
            retinaface_model.build_model(),
            input_signature=(tf.TensorSpec(shape=[None, None, None, 3], dtype=np.float32),)
        )

    return model

def get_image(img_path):
    return img_path

def detect_faces(img_path, threshold=0.9, model = None, allow_upscaling = True):
    """
    TODO: add function doc here
    """

    img = get_image(img_path)

    #---------------------------

    if model is None:
        model = build_model()

    #---------------------------

    nms_threshold = 0.4; decay4=0.5

    _feat_stride_fpn = [32, 16, 8]

    _anchors_fpn = {
        'stride32': np.array([[-248., -248.,  263.,  263.], [-120., -120.,  135.,  135.]], dtype=np.float32),
        'stride16': np.array([[-56., -56.,  71.,  71.], [-24., -24.,  39.,  39.]], dtype=np.float32),
        'stride8': np.array([[-8., -8., 23., 23.], [ 0.,  0., 15., 15.]], dtype=np.float32)
    }

    _num_anchors = {'stride32': 2, 'stride16': 2, 'stride8': 2}

    #---------------------------

    proposals_list = []
    scores_list = []
    landmarks_list = []
    im_tensor, im_info, im_scale = preprocess.preprocess_image(img, allow_upscaling)
    net_out = model(im_tensor)
    net_out = [elt.numpy() for elt in net_out]
    sym_idx = 0

    for _idx, s in enumerate(_feat_stride_fpn):
        _key = 'stride%s'%s
        scores = net_out[sym_idx]
        scores = scores[:, :, :, _num_anchors['stride%s'%s]:]

        bbox_deltas = net_out[sym_idx + 1]
        height, width = bbox_deltas.shape[1], bbox_deltas.shape[2]

        A = _num_anchors['stride%s'%s]
        K = height * width
        anchors_fpn = _anchors_fpn['stride%s'%s]
        anchors = postprocess.anchors_plane(height, width, s, anchors_fpn)
        anchors = anchors.reshape((K * A, 4))
        scores = scores.reshape((-1, 1))

        bbox_stds = [1.0, 1.0, 1.0, 1.0]
        bbox_deltas = bbox_deltas
        bbox_pred_len = bbox_deltas.shape[3]//A
        bbox_deltas = bbox_deltas.reshape((-1, bbox_pred_len))
        bbox_deltas[:, 0::4] = bbox_deltas[:,0::4] * bbox_stds[0]
        bbox_deltas[:, 1::4] = bbox_deltas[:,1::4] * bbox_stds[1]
        bbox_deltas[:, 2::4] = bbox_deltas[:,2::4] * bbox_stds[2]
        bbox_deltas[:, 3::4] = bbox_deltas[:,3::4] * bbox_stds[3]
        proposals = postprocess.bbox_pred(anchors, bbox_deltas)

        proposals = postprocess.clip_boxes(proposals, im_info[:2])

        if s==4 and decay4<1.0:
            scores *= decay4

        scores_ravel = scores.ravel()
        order = np.where(scores_ravel>=threshold)[0]
        proposals = proposals[order, :]
        scores = scores[order]

        proposals[:, 0:4] /= im_scale
        proposals_list.append(proposals)
        scores_list.append(scores)

        landmark_deltas = net_out[sym_idx + 2]
        landmark_pred_len = landmark_deltas.shape[3]//A
        landmark_deltas = landmark_deltas.reshape((-1, 5, landmark_pred_len//5))
        landmarks = postprocess.landmark_pred(anchors, landmark_deltas)
        landmarks = landmarks[order, :]

        landmarks[:, :, 0:2] /= im_scale
        landmarks_list.append(landmarks)
        sym_idx += 3

    proposals = np.vstack(proposals_list)
    if proposals.shape[0]==0:
        landmarks = np.zeros( (0,5,2) )
        return np.zeros( (0,5) ), landmarks
    scores = np.vstack(scores_list)
    scores_ravel = scores.ravel()
    order = scores_ravel.argsort()[::-1]

    proposals = proposals[order, :]
    scores = scores[order]
    landmarks = np.vstack(landmarks_list)
    landmarks = landmarks[order].astype(np.float32, copy=False)

    pre_det = np.hstack((proposals[:,0:4], scores)).astype(np.float32, copy=False)

    #nms = cpu_nms_wrapper(nms_threshold)
    #keep = nms(pre_det)
    keep = postprocess.cpu_nms(pre_det, nms_threshold)

    det = np.hstack( (pre_det, proposals[:,4:]) )
    det = det[keep, :]
    landmarks = landmarks[keep]

    resp = {}
    for idx, face in enumerate(det):

        label = 'face_'+str(idx+1)
        resp[label] = {}
        resp[label]["score"] = face[4]

        resp[label]["facial_area"] = list(face[0:4].astype(int))

        resp[label]["landmarks"] = {}
        resp[label]["landmarks"]["right_eye"] = list(landmarks[idx][0])
        resp[label]["landmarks"]["left_eye"] = list(landmarks[idx][1])
        resp[label]["landmarks"]["nose"] = list(landmarks[idx][2])
        resp[label]["landmarks"]["mouth_right"] = list(landmarks[idx][3])
        resp[label]["landmarks"]["mouth_left"] = list(landmarks[idx][4])

    return resp

def extract_faces(img_path, threshold=0.9, model = None, align = True, allow_upscaling = True):

    resp = []

    #---------------------------

    img = get_image(img_path)

    #---------------------------

    obj = detect_faces(img_path = img, threshold = threshold, model = model, allow_upscaling = allow_upscaling)

    if type(obj) == dict:
        for key in obj:
            identity = obj[key]

            facial_area = identity["facial_area"]
            facial_img = img[facial_area[1]: facial_area[3], facial_area[0]: facial_area[2]]

            if align == True:
                landmarks = identity["landmarks"]
                left_eye = landmarks["left_eye"]
                right_eye = landmarks["right_eye"]
                nose = landmarks["nose"]
                mouth_right = landmarks["mouth_right"]
                mouth_left = landmarks["mouth_left"]

                facial_img = postprocess.alignment_procedure(facial_img, right_eye, left_eye, nose)

            resp.append(facial_img[:, :, ::-1])
    #elif type(obj) == tuple:

    return resp