SWE-Vision：500行代码的视觉AI框架，挑战GPT/Gemini极限

# 伪代码示例
def swe_vision_process(image, question):
    # Step 1: 模型观察图片并生成假设
    hypothesis = model.analyze(image, question)
    
    # Step 2: 模型编写验证代码
    verification_code = model.generate_code(hypothesis)
    
    # Step 3: 执行代码获取结果
    result = execute(verification_code, image)
    
    # Step 4: 根据结果调整答案
    final_answer = model.refine(hypothesis, result)
    
    return final_answer

# 模型生成的验证代码
import cv2
import numpy as np

img = cv2.imread('image.jpg')
hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)

# 定义红色范围
lower_red = np.array([0, 100, 100])
upper_red = np.array([10, 255, 255])
mask = cv2.inRange(hsv, lower_red, upper_red)

# 检测圆形
circles = cv2.HoughCircles(mask, cv2.HOUGH_GRADIENT, 1, 20)
count = len(circles[0]) if circles is not None else 0

print(f"检测到 {count} 个红色圆圈")

from PIL import Image
import torch

# 使用目标检测模型定位物体
cat_bbox = detect_object(img, "cat")
sofa_bbox = detect_object(img, "sofa")

# 比较中心点x坐标
cat_center_x = (cat_bbox[0] + cat_bbox[2]) / 2
sofa_center_x = (sofa_bbox[0] + sofa_bbox[2]) / 2

position = "左边" if cat_center_x < sofa_center_x else "右边"
print(f"猫在沙发的{position}")

# 克隆仓库
git clone https://github.com/unipat-ai/swe-vision

# 安装依赖
pip install -r requirements.txt

# 配置模型API
export OPENAI_API_KEY="your-key"

# 运行示例
python examples/count_objects.py --image test.jpg

from swe_vision import VisionAgent

agent = VisionAgent(model="gpt-4")

result = agent.solve(
    image_path="image.jpg",
    question="图中有几个人？",
    max_iterations=3  # 最多验证3次
)

print(result.answer)
print(result.confidence)
print(result.code_used)  # 查看生成的验证代码