yolo v5 onnxruntime与opencv cv2加载部署推理

参考：https://github.com/hpc203/yolov5-lite-onnxruntime
https://github.com/hpc203/yolov5-dnn-cpp-python

1、onnxruntime 加载推理yolo v5 onnx

import cv2
import numpy as np
import argparse
import onnxruntime as ort
import math

class yolov5_lite():
    def __init__(self, model_pb_path, label_path, confThreshold=0.5, nmsThreshold=0.5, objThreshold=0.5):
        so = ort.SessionOptions()
        so.log_severity_level = 3
        self.net = ort.InferenceSession(model_pb_path, so)
        self.classes = list(map(lambda x: x.strip(), open(label_path, 'r').readlines()))
        self.num_classes = len(self.classes)
        anchors = [[10, 13, 16, 30, 33, 23], [30, 61, 62, 45, 59, 119], [116, 90, 156, 198, 373, 326]]
        self.nl = len(anchors)
        self.na = len(anchors[0]) // 2
        self.no = self.num_classes + 5
        self.grid = [np.zeros(1)] * self.nl
        self.stride = np.array([8., 16., 32.])
        self.anchor_grid = np.asarray(anchors, dtype=np.float32).reshape(self.nl, -1, 2)

        self.confThreshold = confThreshold
        self.nmsThreshold = nmsThreshold
        self.objThreshold = objThreshold
        self.input_shape = (self.net.get_inputs()[0].shape[2], self.net.get_inputs()[0].shape[3])
        
    def resize_image(self, srcimg, keep_ratio=True):
        top, left, newh, neww = 0, 0, self.input_shape[0], self.input_shape[1]
        if keep_ratio and srcimg.shape[0] != srcimg.shape[1]:
            hw_scale = srcimg.shape[0] / srcimg.shape[1]
            if hw_scale > 1:
                newh, neww = self.input_shape[0], int(self.input_shape[1] / hw_scale)
                img = cv2.resize(srcimg, (neww, newh), interpolation=cv2.INTER_AREA)
                left = int((self.input_shape[1] - neww) * 0.5)
                img = cv2.copyMakeBorder(img, 0, 0, left, self.input_shape[1] - neww - left, cv2.BORDER_CONSTANT,
                                         value=0)  # add border
            else:
                newh, neww = int(self.input_shape[0] * hw_scale), self.input_shape[1]
                img = cv2.resize(srcimg, (neww, newh), interpolation=cv2.INTER_AREA)
                top = int((self.input_shape[0] - newh) * 0.5)
                img = cv2.copyMakeBorder(img, top, self.input_shape[0] - newh - top, 0, 0, cv2.BORDER_CONSTANT, value=0)
        else:
            img = cv2.resize(srcimg, self.input_shape, interpolation=cv2.INTER_AREA)
        return img, newh, neww, top, left
    def _make_grid(self, nx=20, ny=20):
        xv, yv = np.meshgrid(np.arange(ny), np.arange(nx))
        return np.stack((xv, yv), 2).reshape((-1, 2)).astype(np.float32)

    def postprocess(self, frame, outs, pad_hw):
        newh, neww, padh, padw = pad_hw
        frameHeight = frame.shape[0]
        frameWidth = frame.shape[1]
        ratioh, ratiow = frameHeight / newh, frameWidth / neww
        # Scan through all the bounding boxes output from the network and keep only the
        # ones with high confidence scores. Assign the box's class label as the class with the highest score.
        classIds = []
        confidences = []
        boxes = []
        for detection in outs:
            scores = detection[5:]
            classId = np.argmax(scores)
            confidence = scores[classId]
            if confidence > self.confThreshold and detection[4] > self.objThreshold:
                center_x = int((detection[0] - padw) * ratiow)
                center_y = int((detection[1] - padh) * ratioh)
                width = int(detection[2] * ratiow)
                height = int(detection[3] * ratioh)
                left = int(center_x - width / 2)
                top = int(center_y - height / 2)
                classIds.append(classId)
                confidences.append(float(confidence))
                boxes.append([left, top, width, height])

        # Perform non maximum suppression to eliminate redundant overlapping boxes with
        # lower confidences.
        indices = cv2.dnn.NMSBoxes(boxes, confidences, self.confThreshold, self.nmsThreshold)
        for i in indices:
            i = i[0] if isinstance(i, (tuple,list)) else i
            box = boxes[i]
            left = box[0]
            top = box[1]
            width = box[2]
            height = box[3]
            frame = self.drawPred(frame, classIds[i], confidences[i], left, top, left + width, top + height)
        return frame
    def drawPred(self, frame, classId, conf, left, top, right, bottom):
        # Draw a bounding box.
        cv2.rectangle(frame, (left, top), (right, bottom), (0, 0, 255), thickness=4)

        label = '%.2f' % conf
        label = '%s:%s' % (self.classes[classId], label)

        # Display the label at the top of the bounding box
        labelSize, baseLine = cv2.getTextSize(label, cv2.FONT_HERSHEY_SIMPLEX, 0.5, 1)
        top = max(top, labelSize[1])
        # cv.rectangle(frame, (left, top - round(1.5 * labelSize[1])), (left + round(1.5 * labelSize[0]), top + baseLine), (255,255,255), cv.FILLED)
        cv2.putText(frame, label, (left, top - 10), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), thickness=2)
        return frame
    def detect(self, srcimg):
        img, newh, neww, top, left = self.resize_image(srcimg)
        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
        img = img.astype(np.float32) / 255.0
        blob = np.expand_dims(np.transpose(img, (2, 0, 1)), axis=0)

        outs = self.net.run(None, {self.net.get_inputs()[0].name: blob})[0].squeeze(axis=0)
        row_ind = 0
        for i in range(self.nl):
            h, w = int(self.input_shape[0] / self.stride[i]), int(self.input_shape[1] / self.stride[i])
            length = int(self.na * h * w)
            if self.grid[i].shape[2:4] != (h, w):
                self.grid[i] = self._make_grid(w, h)

            outs[row_ind:row_ind + length, 0:2] = (outs[row_ind:row_ind + length, 0:2] * 2. - 0.5 + np.tile(
                self.grid[i], (self.na, 1))) * int(self.stride[i])
            outs[row_ind:row_ind + length, 2:4] = (outs[row_ind:row_ind + length, 2:4] * 2) ** 2 * np.repeat(
                self.anchor_grid[i], h * w, axis=0)
            row_ind += length
        srcimg = self.postprocess(srcimg, outs, (newh, neww, top, left))
#         cv2.imwrite('result.jpg', srcimg)
        return srcimg

if __name__=='__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--imgpath', type=str, default=r'D:cvyolov5-lite-onnxruntime-mainimgsbus.jpg', help="image path")
    parser.add_argument('--modelpath', type=str, default=r'D:cvyolov5-lite-onnxruntime-mainonnxmodelv5lite-g.onnx', help="onnx filepath")
    parser.add_argument('--classfile', type=str, default='D:cvyolov5-lite-onnxruntime-mainimgscoco.names', help="classname filepath")
    parser.add_argument('--confThreshold', default=0.5, type=float, help='class confidence')
    parser.add_argument('--nmsThreshold', default=0.6, type=float, help='nms iou thresh')
    args = parser.parse_args()

    srcimg = cv2.imread(args.imgpath)
    print(args.imgpath,srcimg)
    net = yolov5_lite(args.modelpath, args.classfile, confThreshold=args.confThreshold, nmsThreshold=args.nmsThreshold)
    print(net)
    srcimg = net.detect(srcimg)

    winName = 'Deep learning object detection in onnxruntime'
    cv2.namedWindow(winName, cv2.WINDOW_NORMAL)
    cv2.imshow(winName, srcimg)
    cv2.waitKey(0)
    cv2.destroyAllWindows()

2、cv2加载部署推理yolo v5 onnx

import cv2
import argparse
import numpy as np

class yolov5():
    def __init__(self, yolo_type, confThreshold=0.5, nmsThreshold=0.5, objThreshold=0.5):
        anchors = [[4,5,  8,10,  13,16], [23,29,  43,55,  73,105], [146,217,  231,300,  335,433]]
        num_classes = 1
        self.nl = len(anchors)
        self.na = len(anchors[0]) // 2
        self.no = num_classes + 5 + 10
        self.grid = [np.zeros(1)] * self.nl
        self.stride = np.array([8., 16., 32.])
        self.anchor_grid = np.asarray(anchors, dtype=np.float32).reshape(self.nl, -1, 2)
        self.inpWidth = 640
        self.inpHeight = 640
        self.net = cv2.dnn.readNet(yolo_type+'-face.onnx')
        self.confThreshold = confThreshold
        self.nmsThreshold = nmsThreshold
        self.objThreshold = objThreshold

    def _make_grid(self, nx=20, ny=20):
        xv, yv = np.meshgrid(np.arange(ny), np.arange(nx))
        return np.stack((xv, yv), 2).reshape((-1, 2)).astype(np.float32)

    def postprocess(self, frame, outs):
        frameHeight = frame.shape[0]
        frameWidth = frame.shape[1]
        ratioh, ratiow = frameHeight / self.inpHeight, frameWidth / self.inpWidth
        # Scan through all the bounding boxes output from the network and keep only the
        # ones with high confidence scores. Assign the box's class label as the class with the highest score.

        confidences = []
        boxes = []
        landmarks = []
        for detection in outs:
            confidence = detection[15]
            # if confidence > self.confThreshold and detection[4] > self.objThreshold:
            if detection[4] > self.objThreshold:
                center_x = int(detection[0] * ratiow)
                center_y = int(detection[1] * ratioh)
                width = int(detection[2] * ratiow)
                height = int(detection[3] * ratioh)
                left = int(center_x - width / 2)
                top = int(center_y - height / 2)

                confidences.append(float(confidence))
                boxes.append([left, top, width, height])
                landmark = detection[5:15] * np.tile(np.float32([ratiow,ratioh]), 5)
                landmarks.append(landmark.astype(np.int32))
        # Perform non maximum suppression to eliminate redundant overlapping boxes with
        # lower confidences.
        indices = cv2.dnn.NMSBoxes(boxes, confidences, self.confThreshold, self.nmsThreshold)
        print(indices)
        for i in indices:
            # i = i
            box = boxes[i]
            left = box[0]
            top = box[1]
            width = box[2]
            height = box[3]
            landmark = landmarks[i]
            frame = self.drawPred(frame, confidences[i], left, top, left + width, top + height, landmark)
        return frame
    def drawPred(self, frame, conf, left, top, right, bottom, landmark):
        # Draw a bounding box.
        cv2.rectangle(frame, (left, top), (right, bottom), (0, 0, 255), thickness=2)
        # label = '%.2f' % conf
        # Display the label at the top of the bounding box
        # labelSize, baseLine = cv2.getTextSize(label, cv2.FONT_HERSHEY_SIMPLEX, 0.5, 1)
        # top = max(top, labelSize[1])
        # cv2.putText(frame, label, (left, top - 10), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), thickness=2)
        for i in range(5):
            cv2.circle(frame, (landmark[i*2], landmark[i*2+1]), 1, (0,255,0), thickness=-1)
        return frame
    def detect(self, srcimg):
        blob = cv2.dnn.blobFromImage(srcimg, 1 / 255.0, (self.inpWidth, self.inpHeight), [0, 0, 0], swapRB=True, crop=False)
        # Sets the input to the network
        self.net.setInput(blob)

        # Runs the forward pass to get output of the output layers
        outs = self.net.forward(self.net.getUnconnectedOutLayersNames())[0]

        # inference output
        outs[..., [0,1,2,3,4,15]] = 1 / (1 + np.exp(-outs[..., [0,1,2,3,4,15]]))   ###sigmoid
        row_ind = 0
        for i in range(self.nl):
            h, w = int(self.inpHeight/self.stride[i]), int(self.inpWidth/self.stride[i])
            length = int(self.na * h * w)
            if self.grid[i].shape[2:4] != (h,w):
                self.grid[i] = self._make_grid(w, h)
            
            g_i = np.tile(self.grid[i], (self.na, 1))
            a_g_i = np.repeat(self.anchor_grid[i], h * w, axis=0)
            outs[row_ind:row_ind + length, 0:2] = (outs[row_ind:row_ind + length, 0:2] * 2. - 0.5 + g_i) * int(self.stride[i])
            outs[row_ind:row_ind + length, 2:4] = (outs[row_ind:row_ind + length, 2:4] * 2) ** 2 * a_g_i

            outs[row_ind:row_ind + length, 5:7] = outs[row_ind:row_ind + length, 5:7] * a_g_i + g_i * int(self.stride[i])   # landmark x1 y1
            outs[row_ind:row_ind + length, 7:9] = outs[row_ind:row_ind + length, 7:9] * a_g_i + g_i * int(self.stride[i])  # landmark x2 y2
            outs[row_ind:row_ind + length, 9:11] = outs[row_ind:row_ind + length, 9:11] * a_g_i + g_i * int(self.stride[i])  # landmark x3 y3
            outs[row_ind:row_ind + length, 11:13] = outs[row_ind:row_ind + length, 11:13] * a_g_i + g_i * int(self.stride[i])  # landmark x4 y4
            outs[row_ind:row_ind + length, 13:15] = outs[row_ind:row_ind + length, 13:15] * a_g_i + g_i * int(self.stride[i])  # landmark x5 y5
            row_ind += length
        return outs

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument('--yolo_type', type=str, default='yolov5m', choices=['yolov5s', 'yolov5m', 'yolov5l'], help="yolo type")
    parser.add_argument("--imgpath", type=str, default='bus.jpg', help="image path")
    parser.add_argument('--confThreshold', default=0.3, type=float, help='class confidence')
    parser.add_argument('--nmsThreshold', default=0.5, type=float, help='nms iou thresh')
    parser.add_argument('--objThreshold', default=0.3, type=float, help='object confidence')
    args = parser.parse_args()

    yolonet = yolov5(args.yolo_type, confThreshold=args.confThreshold, nmsThreshold=args.nmsThreshold, objThreshold=args.objThreshold)
    srcimg = cv2.imread(args.imgpath)
    dets = yolonet.detect(srcimg)
    srcimg = yolonet.postprocess(srcimg, dets)

    winName = 'Deep learning object detection in OpenCV'
    cv2.namedWindow(winName, 0)
    cv2.imshow(winName, srcimg)
    cv2.waitKey(0)
    cv2.destroyAllWindows()