在OpenVINO Runtime运行YOLOv8输出形状异常，求检测框提取方案

问题 #

我用官方的yolov8n.pt作为基础模型训练了目标检测模型，但在Intel CPU上推理速度太慢，没法落地。于是用Intel的OpenVINO Runtime优化，通过mo命令把ONNX模型转成IR文件后加载运行，但转换后模型输出形状不符合预期，对比yolov8n.yaml和IR文件也搞不懂输出含义。换了yolov8s.pt和不同数据集后，OpenVINO优化后的输出形状还是固定的：

• 输入形状：[1,3,640,640]
• 输出形状：[1,6,8400]

运行代码如下：

import cv2
import matplotlib.pyplot as plt
import numpy as np
from openvino.runtime import Core
from ultralytics import YOLO
ie = Core()

model = ie.read_model(model=r"\models\IR\model_fit.xml")
compiled_model = ie.compile_model(model=model, device_name="CPU")

input_layer_ir = compiled_model.input(0)
output_layer_ir = compiled_model.output()


image = cv2.imread(r'\Resources\test_image.jpg')
N, C, H, W = input_layer_ir.shape
resized_image = cv2.resize(image, (W, H))
input_image = np.expand_dims(resized_image.transpose(2, 0, 1), 0)
plt.imshow(cv2.cvtColor(image, cv2.COLOR_BGR2RGB));
boxes = compiled_model([input_image])[output_layer_ir]

print(boxes.shape)

输出结果：

<Output: names[output0] shape[1,6,8400] type: f32>

我是计算机视觉新手，该怎么处理才能获取检测框坐标？

解决方案 #

1. 先理解输出形状的含义 #

[1,6,8400]的维度定义：

• 1：批量大小（当前为单张输入图）
• 6：每个预测框的6个参数，顺序为 x中心坐标、y中心坐标、框宽度、框高度、目标置信度、类别ID
• 8400：YOLOv8固定的锚框预测总数，不同模型结构下该值不变

2. 解析输出并处理检测框 #

步骤1：调整输出张量形状

先把输出转为更易处理的维度格式：

# 去掉批量维度，转置后得到(8400,6)的形状
predictions = boxes[0].transpose(1, 0)

步骤2：过滤低置信度预测框

设置置信度阈值（比如0.5），筛掉置信度不足的无效框：

conf_threshold = 0.5
valid_predictions = predictions[predictions[:, 4] > conf_threshold]

步骤3：将归一化坐标转成原图像素坐标

YOLOv8输出的坐标是相对于640×640输入图的归一化值，需要先转成输入图的像素坐标，再映射回原图尺寸：

orig_h, orig_w = image.shape[:2]
input_w, input_h = W, H
# 计算输入图到原图的缩放比例
scale_x = orig_w / input_w
scale_y = orig_h / input_h

processed_boxes = []
for pred in valid_predictions:
    x_center, y_center, w, h, conf, cls_id = pred
    # 把归一化坐标转为输入图的像素坐标，再映射到原图
    x1 = int((x_center - w/2) * input_w * scale_x)
    y1 = int((y_center - h/2) * input_h * scale_y)
    x2 = int((x_center + w/2) * input_w * scale_x)
    y2 = int((y_center + h/2) * input_h * scale_y)
    processed_boxes.append([x1, y1, x2, y2, conf, cls_id])

步骤4：非极大值抑制（NMS）去除重复框

同一目标可能被多个框重复检测，用NMS去掉高重叠的冗余框：

nms_threshold = 0.5
if processed_boxes:
    boxes_np = np.array([b[:4] for b in processed_boxes])
    confidences = np.array([b[4] for b in processed_boxes])
    # 调用OpenCV的NMS函数
    indices = cv2.dnn.NMSBoxes(boxes_np.tolist(), confidences.tolist(), conf_threshold, nms_threshold)
    # 得到最终有效检测框
    final_boxes = [processed_boxes[i] for i in indices.flatten()]

步骤5：完整可运行代码

整合所有步骤，最终可以在原图上绘制检测框：

import cv2
import matplotlib.pyplot as plt
import numpy as np
from openvino.runtime import Core

ie = Core()

model = ie.read_model(model=r"\models\IR\model_fit.xml")
compiled_model = ie.compile_model(model=model, device_name="CPU")

input_layer_ir = compiled_model.input(0)
output_layer_ir = compiled_model.output()

image = cv2.imread(r'\Resources\test_image.jpg')
N, C, H, W = input_layer_ir.shape
resized_image = cv2.resize(image, (W, H))
input_image = np.expand_dims(resized_image.transpose(2, 0, 1), 0)

# 推理获取输出
boxes = compiled_model([input_image])[output_layer_ir]

# 调整输出形状
predictions = boxes[0].transpose(1, 0)

# 过滤低置信度框
conf_threshold = 0.5
valid_predictions = predictions[predictions[:, 4] > conf_threshold]

# 坐标转换与映射
orig_h, orig_w = image.shape[:2]
input_w, input_h = W, H
scale_x = orig_w / input_w
scale_y = orig_h / input_h

processed_boxes = []
for pred in valid_predictions:
    x_center, y_center, w, h, conf, cls_id = pred
    x1 = int((x_center - w/2) * input_w * scale_x)
    y1 = int((y_center - h/2) * input_h * scale_y)
    x2 = int((x_center + w/2) * input_w * scale_x)
    y2 = int((y_center + h/2) * input_h * scale_y)
    processed_boxes.append([x1, y1, x2, y2, conf, cls_id])

# NMS去重
nms_threshold = 0.5
if processed_boxes:
    boxes_np = np.array([b[:4] for b in processed_boxes])
    confidences = np.array([b[4] for b in processed_boxes])
    indices = cv2.dnn.NMSBoxes(boxes_np.tolist(), confidences.tolist(), conf_threshold, nms_threshold)
    final_boxes = [processed_boxes[i] for i in indices.flatten()]

    # 在原图绘制检测框与标签
    for box in final_boxes:
        x1, y1, x2, y2, conf, cls_id = box
        cv2.rectangle(image, (x1, y1), (x2, y2), (0, 255, 0), 2)
        label = f"Class {int(cls_id)}: {conf:.2f}"
        cv2.putText(image, label, (x1, y1-10), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0,255,0), 2)

# 显示结果
plt.imshow(cv2.cvtColor(image, cv2.COLOR_BGR2RGB))
plt.show()

内容的提问来源于stack exchange，提问作者wizer_102