""" 输入：原图，图中face框 输出：每张face的5点特征点的位置 """ # from common.face_landmark_points_util import shape_index_process # from common.face_detector_util import prior_box_forward, decode, nms_sorted import cv2 import onnxruntime as ort import numpy as np # 设置ONNX Runtime的日志级别为ERROR ort.set_default_logger_severity(3) # 3表示ERROR级别 m_origin_patch = [15, 15] m_origin = [112, 112] class HypeShape: def __init__(self, shape): self.m_shape = shape self.m_weights = [0]*len(self.m_shape) size = len(self.m_shape) self.m_weights[size - 1] = self.m_shape[size - 1] for times in range(size - 1): self.m_weights[size - 1 - times - 1] = self.m_weights[size - 1 - times] * self.m_shape[size - 1 - times - 1] def to_index(self, coordinate): if len(coordinate) == 0: return 0 size = len(coordinate) weight_start = len(self.m_weights) - size + 1 index = 0 for times in range(size - 1): index += self.m_weights[weight_start + times] * coordinate[times] index += coordinate[size - 1] return index def shape_index_process(feat_data, pos_data): feat_h = feat_data.shape[2] feat_w = feat_data.shape[3] landmarkx2 = pos_data.shape[1] x_patch_h = int( m_origin_patch[0] * feat_data.shape[2] / float( m_origin[0] ) + 0.5 ) x_patch_w = int( m_origin_patch[1] * feat_data.shape[3] / float( m_origin[1] ) + 0.5 ) feat_patch_h = x_patch_h feat_patch_w = x_patch_w num = feat_data.shape[0] channels = feat_data.shape[1] r_h = ( feat_patch_h - 1 ) / 2.0 r_w = ( feat_patch_w - 1 ) / 2.0 landmark_num = int(landmarkx2 * 0.5) pos_offset = HypeShape([pos_data.shape[0], pos_data.shape[1]]) feat_offset = HypeShape([feat_data.shape[0], feat_data.shape[1], feat_data.shape[2], feat_data.shape[3]]) nmarks = int( landmarkx2 * 0.5 ) out_shape = [feat_data.shape[0], feat_data.shape[1], x_patch_h, nmarks, x_patch_w] out_offset = HypeShape([feat_data.shape[0], feat_data.shape[1], x_patch_h, nmarks, x_patch_w]) buff = np.zeros(out_shape) zero = 0 buff = buff.reshape((-1)) pos_data = pos_data.reshape((-1)) feat_data = feat_data.reshape((-1)) for i in range(landmark_num): for n in range(num): # coordinate of the first patch pixel, scale to the feature map coordinate y = int( pos_data[pos_offset.to_index( [n, 2 * i + 1] )] * ( feat_h - 1 ) - r_h + 0.5 ) x = int( pos_data[pos_offset.to_index( [n, 2 * i] )] * ( feat_w - 1 ) - r_w + 0.5 ) for c in range(channels): for ph in range(feat_patch_h): for pw in range(feat_patch_w): y_p = y + ph x_p = x + pw # set zero if exceed the img bound if y_p < 0 or y_p >= feat_h or x_p < 0 or x_p >= feat_w: buff[out_offset.to_index( [n, c, ph, i, pw] )] = zero else: buff[out_offset.to_index( [n, c, ph, i, pw] )] = feat_data[feat_offset.to_index( [n, c, y_p, x_p] )] return buff.reshape((1,-1,1,1)).astype(np.float32) def crop_face(image:np.ndarray, face, H, W): """ Crop a face from an image with padding if the face is out of bounds. Args: image (np.ndarray): The input image as a numpy array of shape (H, W, C). face (tuple): A tuple containing (x, y, w, h) for the face rectangle. padding (list): Padding data for padding the image. Returns: np.ndarray: Cropped and padded image. """ x0, y0, x1, y1 = int(round(face[0])), int(round(face[1])), int(round(face[2])), int(round(face[3])) # Calculate padding pad_top = max(0, -y0) pad_bottom = max(0, y1 - H) pad_left = max(0, -x0) pad_right = max(0, x1 - W) # Apply padding padded_image = np.pad(image, ((pad_top, pad_bottom), (pad_left, pad_right), (0, 0)), mode='constant', constant_values=0) # Update new coordinates after padding new_x0, new_y0 = x0 + pad_left, y0 + pad_top new_x1, new_y1 = x1 + pad_left, y1 + pad_top # Crop the face cropped_image = padded_image[new_y0:new_y1, new_x0:new_x1, :] return cropped_image, (x0,y0,x1,y1) class Landmark5er(): def __init__(self, onnx_path1, onnx_path2, num_threads=1) -> None: session_options = ort.SessionOptions() session_options.intra_op_num_threads = num_threads # 初始化 InferenceSession 时传入 SessionOptions 对象 self.ort_session1 = ort.InferenceSession(onnx_path1, session_options=session_options) self.ort_session2 = ort.InferenceSession(onnx_path2, session_options=session_options) self.first_input_name = self.ort_session1.get_inputs()[0].name self.second_input_name = self.ort_session2.get_inputs()[0].name def inference(self, image, box): # face_img = image[int(box[1]):int(box[3]), int(box[0]):int(box[2])] # 这种裁剪不对,不合适 H,W,C = image.shape face_img, box = crop_face(image,box,H,W) x1,y1,x2,y2 = int(box[0]) ,int(box[1]), int(box[2]),int(box[3]) if x1 < 0 or x1 > W-1 or x2 < 0 or x2 > W: print("x超出边界") if y1 < 0 or y1 > H-1 or y2 < 0 or y2 > H: print("y超出边界") face_img = cv2.resize(face_img,(112,112)) gray_img = cv2.cvtColor(face_img, cv2.COLOR_BGR2GRAY) gray_img = gray_img.reshape((1, 1, 112, 112)).astype(np.float32) # 输入必须是(1,1,112,112) # points5 net1 results_1 = self.ort_session1.run([], {self.first_input_name: gray_img}) # shape index process feat_data = results_1[0] pos_data = results_1[1] shape_index_results = shape_index_process(feat_data, pos_data) results_2 = self.ort_session2.run([], {self.second_input_name: shape_index_results}) landmarks = (results_2[0] + results_1[1]) * 112 # print("results_2[0] , results_1[1]: ",results_2[0], results_1[1]) # print("in find_landmarks, landmarks:", landmarks) landmarks = landmarks.reshape((-1)).astype(np.int32) scale_x = (box[2] - box[0]) / 112.0 scale_y = (box[3] - box[1]) / 112.0 mapped_landmarks = [] for i in range(landmarks.size // 2): x = box[0] + landmarks[2 * i] * scale_x y = box[1] + landmarks[2 * i + 1] * scale_y x = max(0.01, min(x,W-0.01)) y = max(0.01, min(y,H-0.01)) mapped_landmarks.append((x, y)) return mapped_landmarks if __name__ == "__main__": import sys ld5 = Landmark5er(onnx_path1="./checkpoints/face_landmarker_pts5_net1.onnx", onnx_path2="./checkpoints/face_landmarker_pts5_net2.onnx", num_threads=1) if len(sys.argv) > 1: jpg_path = sys.argv[1] else: jpg_path = "asserts/1.jpg" image = cv2.imread(jpg_path) if image is None: print("Error: Could not load image.") exit() box = (201.633087308643, 42.78490193881931, 319.49375572419393, 191.68867463550623) landmarks5 = ld5.inference(image,box) print(landmarks5)