YOLOv6人脸Landmark检测

官方代码与参考文档给出的解析顺序，xyxy, conf, cls, lmdks，这部分还有官方的参考文件：

https://github.com/meituan/YOLOv6/blob/yolov6-face/yolov6/core/inferer.py

def wrap_detection(self, input_image, out_data):
    confidences = []
    boxes = []
    kypts = []
    rows = out_data.shape[0]

    image_width, image_height, _ = input_image.shape

    x_factor = image_width / 640.0
    y_factor = image_height / 640.0

    sd = np.zeros((5, 2), dtype=np.float32)
    sd[0:5] = (x_factor, y_factor)
    sd = np.squeeze(sd.reshape((-1, 1)), 1)
    # xyxy, lmdks, conf, cls,
    for r in range(rows):
        row = out_data[r]
        conf = row[14]
        cls = row[15]
        if (conf > 0.25 and cls > 0.25):
            confidences.append(conf)
            x, y, w, h = row[0].item(), row[1].item(), row[2].item(), row[3].item()
            left = int((x - 0.5 * w) * x_factor)
            top = int((y - 0.5 * h) * y_factor)
            width = int(w * x_factor)
            height = int(h * y_factor)
            box = np.array([left, top, width, height])
            boxes.append(box)
            kypts.append(np.multiply(row[4:14], sd))

    indexes = cv.dnn.NMSBoxes(boxes, confidences, 0.25, 0.25)

    result_confidences = []
    result_boxes = []
    result_kypts = []

    for i in indexes:
        result_confidences.append(confidences[i])
        result_boxes.append(boxes[i])
        result_kypts.append(kypts[i])

    return result_kypts, result_confidences, result_boxes

扫码查看OpenCV+OpenVIO+Pytorch系统化学习路线图