MASt3R （2）安装和测试
模型下载位置
例子1 单独的匹配
from mast3r.model import AsymmetricMASt3R
from mast3r.fast_nn import fast_reciprocal_NNs
import mast3r.utils.path_to_dust3r
from dust3r.inference import inference
from dust3r.utils.image import load_images
import cv2
import numpy as np
import torch
import time
class FeatureMatcher_Mast3r:
def __init__(self, model_path, device='cuda'):
"""
初始化特征匹配器
Args:
model_path: MASt3R模型路径
device: 运行设备
"""
self.device = device
self.model = self.load_mast3r_model(model_path)
def load_mast3r_model(self, model_path):
"""加载MASt3R模型"""
print("开始加载MASt3R模型")
model = AsymmetricMASt3R.from_pretrained(model_path).to(self.device)
print("MASt3R模型加载完成")
return model
def load_and_preprocess_images(self, image_paths, size=512):
"""
加载并预处理图像
Args:
image_paths: 图像路径列表
size: 图像尺寸
Returns:
预处理后的图像数据
"""
images = load_images(image_paths, size=size)
# 图像标准化
image_mean = torch.as_tensor([0.5, 0.5, 0.5], device='cpu').reshape(1, 3, 1, 1)
image_std = torch.as_tensor([0.5, 0.5, 0.5], device='cpu').reshape(1, 3, 1, 1)
viz_imgs = []
for i, view in enumerate([images[0], images[1]]):
rgb_tensor = view['img'] * image_std + image_mean
viz_img = rgb_tensor.squeeze(0).permute(1, 2, 0).cpu().numpy()
viz_imgs.append(viz_img)
# 转换为OpenCV格式
img0_bgr = (viz_imgs[0][:, :, ::-1] * 255).astype(np.uint8)
img1_bgr = (viz_imgs[1][:, :, ::-1] * 255).astype(np.uint8)
return images, img0_bgr, img1_bgr
def mast3r_feature_matching(self, images):
"""
MASt3R特征匹配
Args:
images: 输入图像
Returns:
匹配结果、关键点坐标和耗时
"""
start_time = time.time()
# 执行推理
output = inference([tuple(images)], self.model, self.device, batch_size=1, verbose=False)
view1, pred1 = output['view1'], output['pred1']
view2, pred2 = output['view2'], output['pred2']
desc1, desc2 = pred1['desc'].squeeze(0).detach(), pred2['desc'].squeeze(0).detach()
# 快速互最近邻匹配
matches_im0, matches_im1 = fast_reciprocal_NNs(desc1, desc2, subsample_or_initxy1=8,
device=self.device, dist='dot', block_size=2**13)
# 有效点筛选
H0, W0 = view1['true_shape'][0]
valid_matches_im0 = (matches_im0[:, 0] >= 3) & (matches_im0[:, 0] < int(W0) - 3) & (
matches_im0[:, 1] >= 3) & (matches_im0[:, 1] < int(H0) - 3)
H1, W1 = view2['true_shape'][0]
valid_matches_im1 = (matches_im1[:, 0] >= 3) & (matches_im1[:, 0] < int(W1) - 3) & (
matches_im1[:, 1] >= 3) & (matches_im1[:, 1] < int(H1) - 3)
valid_matches = valid_matches_im0 & valid_matches_im1
matches_im0, matches_im1 = matches_im0[valid_matches], matches_im1[valid_matches]
end_time = time.time()
total_time = end_time - start_time
return matches_im0, matches_im1, total_time
def estimate_pose(self, src_pts, dst_pts, method='homography'):
"""
估计位姿变换（旋转矩阵R和平移向量t）
Args:
src_pts: 源点
dst_pts: 目标点
method: 估计方法
Returns:
R, t, 内点数量
"""
if len(src_pts) < 4:
print("匹配点数量不足，无法计算位姿")
return None, None, 0
if method == 'homography':
# 使用单应性矩阵估计
H, mask = cv2.findHomography(src_pts, dst_pts, cv2.RANSAC, 5.0)
if H is not None:
num_inliers = np.sum(mask)
return H, None, num_inliers
else:
return None, None, 0
elif method == 'essential':
# 使用本质矩阵估计（需要相机内参）
camera_matrix = np.array([[1000, 0, 512],
[0, 1000, 512],
[0, 0, 1]])
E, mask = cv2.findEssentialMat(src_pts, dst_pts, camera_matrix, cv2.RANSAC, 0.999, 1.0)
if E is not None:
points, R, t, mask = cv2.recoverPose(E, src_pts, dst_pts, camera_matrix)
num_inliers = np.sum(mask)
return R, t, num_inliers
else:
return None, None, 0
return None, None, 0
def visualize_results(self, mast3r_matches_im0, mast3r_matches_im1, img0_bgr, img1_bgr, n_viz=100):
"""
可视化MASt3R匹配结果
"""
h0, w0 = img0_bgr.shape[:2]
h1, w1 = img1_bgr.shape[:2]
max_height = max(h0, h1)
# 创建结果图像
result_img = np.zeros((max_height, w0 + w1, 3), dtype=np.uint8)
# 放置原始图像
result_img[:h0, :w0] = img0_bgr
result_img[:h1, w0:w0+w1] = img1_bgr
# 绘制MASt3R匹配点
colors = [
(255, 0, 0), (0, 255, 0), (0, 0, 255), (255, 255, 0), (255, 0, 255)
]
n_viz_actual = min(n_viz, mast3r_matches_im0.shape[0])
indices = np.round(np.linspace(0, mast3r_matches_im0.shape[0] - 1, n_viz_actual)).astype(int)
for i, idx in enumerate(indices):
color = colors[i % len(colors)]
pt1 = mast3r_matches_im0[idx]
pt2 = mast3r_matches_im1[idx]
x1, y1 = int(pt1[0]), int(pt1[1])
x2, y2 = int(pt2[0]) + w0, int(pt2[1])
cv2.line(result_img, (x1, y1), (x2, y2), color, 1, cv2.LINE_AA)
cv2.circle(result_img, (x1, y1), 3, color, -1)
cv2.circle(result_img, (x2, y2), 3, color, -1)
# 添加标签
cv2.putText(result_img, "MASt3R Feature Matching", (10, 30),
cv2.FONT_HERSHEY_SIMPLEX, 0.7, (255, 255, 255), 2)
return result_img
def main():
"""主函数"""
# 初始化参数
model_path = "checkpoints/MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric.pth"
image_paths = [
'/media/r9000k/DD_XS/2数据/2_1精灵4无rtk/2圆楼_滨湖_白天_原图/DJI_0516.JPG',
'/media/r9000k/DD_XS/2数据/2_1精灵4无rtk/v4_2圆楼_滨湖_拍照模式测曝光/DJI_0146.JPG'
]
# 创建特征匹配器实例
matcher_Mast3r = FeatureMatcher_Mast3r(model_path)
# 加载图像
images, img0_bgr, img1_bgr = matcher_Mast3r.load_and_preprocess_images(image_paths)
# MASt3R特征匹配
mast3r_pts1, mast3r_pts2, mast3r_time = matcher_Mast3r.mast3r_feature_matching(images)
# 位姿估计
print("\n=== MASt3R位姿估计结果 ===")
R_mast3r, t_mast3r, mast3r_inliers = matcher_Mast3r.estimate_pose(mast3r_pts1, mast3r_pts2)
print(f"MASt3R内点数量: {mast3r_inliers}")
print(f"MASt3R算法耗时: {mast3r_time:.4f} 秒")
print(f"匹配点数量: {mast3r_pts1.shape[0]}")
# 可视化结果
result_img = matcher_Mast3r.visualize_results(mast3r_pts1, mast3r_pts2, img0_bgr, img1_bgr, n_viz=20)
# 显示结果
cv2.namedWindow('MASt3R Feature Matching Results', cv2.WINDOW_NORMAL)
cv2.imshow('MASt3R Feature Matching Results', result_img)
cv2.waitKey(0)
cv2.destroyAllWindows()
if __name__ == '__main__':
main()
效果对比
原始版本
https://github.com/naver/mast3r
加速版本 必须cuda12.1 装不上
peedy MASt3R 在不影响准确率的情况下，将推理时间缩短了 54%（每对图像从 198 毫秒缩短至 91 毫秒）。这一改进实现实时 3D 理解，助力MR导航和大规模 3D 场景重建等应用。
https://github.com/ASU-ESIC-FAN-Lab/speedy_mast3r
- 克隆 MASt3R。
git clone --recursive https://github.com/naver/mast3r
cd mast3r
git submodule update --init --recursive
- 创建环境，这里我们以 conda 为例进行说明。 注意适配自己的环境
conda create -n mast3r python=3.11 cmake=3.14.0 conda activate mast3r conda install pytorch=2.7 torchvision pytorch-cuda=11.8 -c pytorch -c nvidia # use the correct version of cuda for your system pip install -r requirements.txt pip install -r dust3r/requirements.txt # Optional: you can also install additional packages to: # - add support for HEIC images # - add required packages for visloc.py pip install -r dust3r/requirements_optional.txt
- 编译并安装 ASMK
pip install cython git clone https://github.com/jenicek/asmk cd asmk/cython/ cythonize *.pyx cd ..
pip install backports.tarfile
conda update setuptools pip backports
conda install -c conda-forge libgcc-ng libstdcxx-ng
sudo apt-get update
sudo apt-get install libc6-dev
python3 setup.py build_ext --inplace
或者
pip install . cd ..
1报错libpthread找不到
查找系统库文件位置：首先需要确认您系统的 libpthread.so.0文件在哪里。
sudo find / -name libpthread.so.0 2>/dev/null
/home/r9000k/anaconda3/pkgs/sysroot_linux-64-2.17-h0157908_18/x86_64-conda-linux-gnu/sysroot/lib64/libpthread.so.0 /home/r9000k/anaconda3/envs/pyslam/x86_64-conda-linux-gnu/sysroot/lib64/libpthread.so.0 /usr/lib/i386-linux-gnu/libpthread.so.0 /usr/lib/x86_64-linux-gnu/libpthread.so.0 /usr/lib32/libpthread.so.0 /opt/deepinwine/runtime-i386/lib/i386-linux-gnu/libpthread.so.0 /opt/deepinwine/runtime-i386/lib/x86_64-linux-gnu/libpthread.so.0 /snap/snapd/25202/usr/lib/x86_64-linux-gnu/libpthread.so.0 /snap/snapd/25577/usr/lib/x86_64-linux-gnu/libpthread.so.0 /snap/core20/2686/usr/lib/i386-linux-gnu/libpthread.so.0 /snap/core20/2686/usr/lib/x86_64-linux-gnu/libpthread.so.0 /snap/core20/2682/usr/lib/i386-linux-gnu/libpthread.so.0 /snap/core20/2682/usr/lib/x86_64-linux-gnu/libpthread.so.0 /snap/core22/2163/usr/lib/i386-linux-gnu/libpthread.so.0 /snap/core22/2163/usr/lib/x86_64-linux-gnu/libpthread.so.0 /snap/core22/2193/usr/lib/i386-linux-gnu/libpthread.so.0 /snap/core22/2193/usr/lib/x86_64-linux-gnu/libpthread.so.0 /snap/core18/2976/lib/i386-linux-gnu/libpthread.so.0 /snap/core18/2976/lib/x86_64-linux-gnu/libpthread.so.0 /snap/core18/2959/lib/i386-linux-gnu/libpthread.so.0 /snap/core18/2959/lib/x86_64-linux-gnu/libpthread.so.0 /snap/core24/1237/usr/lib/i386-linux-gnu/libpthread.so.0 /snap/core24/1237/usr/lib/x86_64-linux-gnu/libpthread.so.0 /snap/core24/1243/usr/lib/i386-linux-gnu/libpthread.so.0 /snap/core24/1243/usr/lib/x86_64-linux-gnu/libpthread.so.0
查找你的Conda环境中的系统库路径。通常它在 $CONDA_PREFIX/x86_64-conda-linux-gnu/sysroot/下。你可以使用以下命令查找：
find $CONDA_PREFIX -name "libpthread.so.0" 2>/dev/null
假设找到了相关文件，使用 cp -inr命令将Conda环境中的库文件复制到系统库目录（注意：使用 -n参数避免覆盖已有文件）
sudo cp -inr $CONDA_PREFIX/x86_64-conda-linux-gnu/sysroot/lib64/* /lib64/ sudo cp -inr $CONDA_PREFIX/x86_64-conda-linux-gnu/sysroot/usr/lib64/* /usr/lib64/
成功新安装看起来你的编译过程已经成功了！
报错2
（可选）编译 RoPE 的 cuda 内核（如 CroCo v2 中那样）先跳过
这个错误是因为你的 CUDA 11.8 与当前使用的 GCC 编译器版本不兼容。错误信息明确提示：unsupported GNU version! gcc versions later than 11 are not supported!。这意味着 CUDA 11.8 不支持版本高于 11 的 GCC 编译器。
conda install -c conda-forge gxx_linux-64=11.*
下载权重
mkdir -p checkpoints/ wget https://download.europe.naverlabs.com/ComputerVision/MASt3R/MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric.pth -P checkpoints/
选择是否用cpu或者cuda
选择模型是在线读取还是离线文件
#device = 'cpu'
device = 'cuda'
#model_name = "naver/MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric"
model_name = "checkpoints/MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric.pth"
print("开始加载MASt3R模型")
model = AsymmetricMASt3R.from_pretrained(model_name).to(device)
print("加载模型完毕，加载图")
测试代码1 双图匹配matplotlib显示
自动下载全红和数据
from mast3r.model import AsymmetricMASt3R
from mast3r.fast_nn import fast_reciprocal_NNs
import mast3r.utils.path_to_dust3r
from dust3r.inference import inference
from dust3r.utils.image import load_images
if __name__ == '__main__':
device = 'cuda'
model_name = "naver/MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric"
# you can put the path to a local checkpoint in model_name if needed
model = AsymmetricMASt3R.from_pretrained(model_name).to(device)
images = load_images(['dust3r/croco/assets/Chateau1.png', 'dust3r/croco/assets/Chateau2.png'], size=512)
output = inference([tuple(images)], model, device, batch_size=1, verbose=False)
# at this stage, you have the raw dust3r predictions
view1, pred1 = output['view1'], output['pred1']
view2, pred2 = output['view2'], output['pred2']
desc1, desc2 = pred1['desc'].squeeze(0).detach(), pred2['desc'].squeeze(0).detach()
# find 2D-2D matches between the two images
matches_im0, matches_im1 = fast_reciprocal_NNs(desc1, desc2, subsample_or_initxy1=8,
device=device, dist='dot', block_size=2**13)
# ignore small border around the edge
H0, W0 = view1['true_shape'][0]
valid_matches_im0 = (matches_im0[:, 0] >= 3) & (matches_im0[:, 0] < int(W0) - 3) & (
matches_im0[:, 1] >= 3) & (matches_im0[:, 1] < int(H0) - 3)
H1, W1 = view2['true_shape'][0]
valid_matches_im1 = (matches_im1[:, 0] >= 3) & (matches_im1[:, 0] < int(W1) - 3) & (
matches_im1[:, 1] >= 3) & (matches_im1[:, 1] < int(H1) - 3)
valid_matches = valid_matches_im0 & valid_matches_im1
matches_im0, matches_im1 = matches_im0[valid_matches], matches_im1[valid_matches]
# visualize a few matches
import numpy as np
import torch
import torchvision.transforms.functional
from matplotlib import pyplot as pl
n_viz = 20
num_matches = matches_im0.shape[0]
match_idx_to_viz = np.round(np.linspace(0, num_matches - 1, n_viz)).astype(int)
viz_matches_im0, viz_matches_im1 = matches_im0[match_idx_to_viz], matches_im1[match_idx_to_viz]
image_mean = torch.as_tensor([0.5, 0.5, 0.5], device='cpu').reshape(1, 3, 1, 1)
image_std = torch.as_tensor([0.5, 0.5, 0.5], device='cpu').reshape(1, 3, 1, 1)
viz_imgs = []
for i, view in enumerate([view1, view2]):
rgb_tensor = view['img'] * image_std + image_mean
viz_imgs.append(rgb_tensor.squeeze(0).permute(1, 2, 0).cpu().numpy())
H0, W0, H1, W1 = *viz_imgs[0].shape[:2], *viz_imgs[1].shape[:2]
img0 = np.pad(viz_imgs[0], ((0, max(H1 - H0, 0)), (0, 0), (0, 0)), 'constant', constant_values=0)
img1 = np.pad(viz_imgs[1], ((0, max(H0 - H1, 0)), (0, 0), (0, 0)), 'constant', constant_values=0)
img = np.concatenate((img0, img1), axis=1)
pl.figure()
pl.imshow(img)
cmap = pl.get_cmap('jet')
for i in range(n_viz):
(x0, y0), (x1, y1) = viz_matches_im0[i].T, viz_matches_im1[i].T
pl.plot([x0, x1 + W0], [y0, y1], '-+', color=cmap(i / (n_viz - 1)), scalex=False, scaley=False)
pl.show(block=True)
报错
这个警告和错误提示表明您的 DUSt3R 项目遇到了两个主要问题：RoPE2D 的 CUDA 内核编译失败，以及 Qt 图形界面库的兼容性问题。下面我来帮您分析和解决。
问题分析
从您的输出信息可以看到两个核心问题：
-
RoPE2D CUDA 内核编译失败：
Warning, cannot find cuda-compiled version of RoPE2D, using a slow pytorch version instead -
Qt 图形界面库加载失败：导致程序崩溃
pip install "numpy<2"
测试代码2 纯opencv显示
默认从云端加载权重
print("开始加载模型")
model = AsymmetricMASt3R.from_pretrained(model_name).to(device)
#model_path = "./checkpoints/MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric"
#model = AsymmetricMASt3R.from_pretrained(model_name).to(device)
from mast3r.model import AsymmetricMASt3R
from mast3r.fast_nn import fast_reciprocal_NNs
import mast3r.utils.path_to_dust3r
from dust3r.inference import inference
from dust3r.utils.image import load_images
import cv2
import numpy as np
import torch
if __name__ == '__main__':
device = 'cuda'
model_name = "naver/MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric"
print("开始加载模型")
model = AsymmetricMASt3R.from_pretrained(model_name).to(device)
#model_path = "./checkpoints/MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric"
#model = AsymmetricMASt3R.from_pretrained(model_name).to(device)
print("加载模型完毕，加载图")
images = load_images(['dust3r/croco/assets/Chateau1.png', 'dust3r/croco/assets/Chateau2.png'], size=512)
output = inference([tuple(images)], model, device, batch_size=1, verbose=False)
view1, pred1 = output['view1'], output['pred1']
view2, pred2 = output['view2'], output['pred2']
desc1, desc2 = pred1['desc'].squeeze(0).detach(), pred2['desc'].squeeze(0).detach()
matches_im0, matches_im1 = fast_reciprocal_NNs(desc1, desc2, subsample_or_initxy1=8,
device=device, dist='dot', block_size=2**13)
print("匹配结束么，开始挑选有效点")
H0, W0 = view1['true_shape'][0]
valid_matches_im0 = (matches_im0[:, 0] >= 3) & (matches_im0[:, 0] < int(W0) - 3) & (
matches_im0[:, 1] >= 3) & (matches_im0[:, 1] < int(H0) - 3)
H1, W1 = view2['true_shape'][0]
valid_matches_im1 = (matches_im1[:, 0] >= 3) & (matches_im1[:, 0] < int(W1) - 3) & (
matches_im1[:, 1] >= 3) & (matches_im1[:, 1] < int(H1) - 3)
valid_matches = valid_matches_im0 & valid_matches_im1
matches_im0, matches_im1 = matches_im0[valid_matches], matches_im1[valid_matches]
# 可视化部分 - 使用OpenCV替代Matplotlib
n_viz = 20
num_matches = matches_im0.shape[0]
match_idx_to_viz = np.round(np.linspace(0, num_matches - 1, n_viz)).astype(int)
viz_matches_im0, viz_matches_im1 = matches_im0[match_idx_to_viz], matches_im1[match_idx_to_viz]
image_mean = torch.as_tensor([0.5, 0.5, 0.5], device='cpu').reshape(1, 3, 1, 1)
image_std = torch.as_tensor([0.5, 0.5, 0.5], device='cpu').reshape(1, 3, 1, 1)
viz_imgs = []
for i, view in enumerate([view1, view2]):
rgb_tensor = view['img'] * image_std + image_mean
viz_imgs.append(rgb_tensor.squeeze(0).permute(1, 2, 0).cpu().numpy())
# OpenCV可视化代码
print('OpenCV可视化代码')
img0_bgr = (viz_imgs[0][:, :, ::-1] * 255).astype(np.uint8)
img1_bgr = (viz_imgs[1][:, :, ::-1] * 255).astype(np.uint8)
h0, w0 = img0_bgr.shape[:2]
h1, w1 = img1_bgr.shape[:2]
max_height = max(h0, h1)
concatenated_img = np.zeros((max_height, w0 + w1, 3), dtype=np.uint8)
concatenated_img[:h0, :w0] = img0_bgr
concatenated_img[:h1, w0:w0+w1] = img1_bgr
colors = [
(255, 0, 0), (0, 255, 0), (0, 0, 255), (255, 255, 0), (255, 0, 255),
(0, 255, 255), (128, 0, 0), (0, 128, 0), (0, 0, 128), (128, 128, 0),
(128, 0, 128), (0, 128, 128), (192, 192, 192), (255, 165, 0), (255, 192, 203),
(138, 43, 226), (165, 42, 42), (188, 143, 143), (95, 158, 160), (100, 149, 237)
]
for i, (pt0, pt1) in enumerate(zip(viz_matches_im0, viz_matches_im1)):
color = colors[i % len(colors)]
x0, y0 = int(pt0[0]), int(pt0[1])
x1, y1 = int(pt1[0]), int(pt1[1])
cv2.line(concatenated_img, (x0, y0), (x1 + w0, y1), color, 2, cv2.LINE_AA)
cv2.circle(concatenated_img, (x0, y0), 5, color, -1)
cv2.circle(concatenated_img, (x1 + w0, y1), 5, color, -1)
cv2.imshow('Feature Matches', concatenated_img)
cv2.waitKey(0)
cv2.destroyAllWindows()
测试3 对比orb 效果
cpu模式
gpu模式 占用3.6G
开始以前
结束以后
优化代码
======================
错误匹配
先放一个正确的
正确匹配
同高度不同时间 不同高度
不同高度
白天和夜晚
120米 光圈2.8 曝光时间1/30秒 增益1600 对比3200
120米 光圈2.8 曝光时间1/30秒 增益3200
120米 光圈2.8 曝光时间1/30秒 增益1600
120米 光圈2.8 曝光时间1/30秒 增益3200
import cv2
import numpy as np
import random
draw_num=20 # 画出来前多少个点
line_thickness_set=4
font_scale = 2.5
color_title = (255, 255, 255) # 蓝 绿 红
thickness = 6
colors_Line = [
(255, 0, 0), (0, 255, 0), (0, 0, 255), (255, 255, 0), (255, 0, 255),(0, 255, 255),(0, 100, 255),(0, 255, 100)
]
def generate_pastel_color():
"""生成柔和的随机颜色"""
r = random.randint(30, 255) # 限制在较亮的范围
g = random.randint(30, 255)
b = random.randint(30, 255)
return (r, g, b)
def draw_ORB_matches(img1, kp1, img2, kp2, top_n_matches, line_thickness=line_thickness_set):
"""
根据匹配质量(distance)动态分配颜色：高质量→绿蓝色，低质量→橘红色
修改版本：直接接受已筛选的top_n_matches列表
参数:
img1: 第一张图像
kp1: 第一张图像的关键点
img2: 第二张图像
kp2: 第二张图像的关键点
top_n_matches: 已经筛选好的匹配列表（通常是按距离排序后的前N个）
line_thickness: 连线粗细
"""
# 获取图像尺寸
h1, w1 = img1.shape[:2]
h2, w2 = img2.shape[:2]
# 创建拼接图像
new_width = w1 + w2
new_height = max(h1, h2)
result = np.zeros((new_height, new_width, 3), dtype=np.uint8)
result[0:h1, 0:w1] = img1
result[0:h2, w1:w1+w2] = img2
# 检查匹配列表是否为空
if len(top_n_matches) == 0:
print("警告：匹配列表为空")
return result
# 获取所有匹配的距离值
distances = [m.distance for m in top_n_matches]
min_dist = min(distances)
max_dist = max(distances)
print(f"绘制 {len(top_n_matches)} 个匹配点")
print(f"最小距离: {min_dist:.2f}, 最大距离: {max_dist:.2f}")
# 绘制每个匹配对
for i, match in enumerate(top_n_matches):
# # 获取关键点坐标
pt1 = (int(kp1[match.queryIdx].pt[0]), int(kp1[match.queryIdx].pt[1]))
pt2 = (int(kp2[match.trainIdx].pt[0] + w1), int(kp2[match.trainIdx].pt[1]))
# # 根据匹配质量计算颜色
# # 归一化距离 (0-1范围，0表示最好，1表示最差)
# # # 根据匹配质量计算颜色（距离越小，匹配质量越高）
# norm_dist = (match.distance - min_dist) / (max_dist - min_dist + 1e-5)
# # 颜色映射：高质量→绿色，中等质量→黄色，低质量→红色
# if norm_dist < 0.33: # 最好的33% - 绿色系
# blue = 0
# green = 255
# red = int(255 * (norm_dist / 0.33))
# elif norm_dist < 0.66: # 中间33% - 黄色到橙色
# blue = 0
# green = int(255 * (1 - (norm_dist - 0.33) / 0.33))
# red = 255
# else: # 最差的33% - 橙色到红色
# blue = 0
# green = int(128 * (1 - (norm_dist - 0.66) / 0.34))
# red = 255
color = colors_Line[i % len(colors_Line)]
#color=generate_pastel_color()
# 绘制连线
cv2.line(result, pt1, pt2, color, line_thickness)
# # 可选：在关键点旁边显示匹配质量和序号
# if line_thickness > 1: # 只在连线较粗时显示文本，避免过于拥挤
# cv2.putText(result, f"{i}", (pt1[0]+5, pt1[1]),
# cv2.FONT_HERSHEY_SIMPLEX, 0.4, (255, 255, 255), 1)
# cv2.putText(result, f"{match.distance:.1f}", (pt2[0]+5, pt2[1]),
# cv2.FONT_HERSHEY_SIMPLEX, 0.4, (255, 255, 255), 1)
return result
# 使用示例
def Orb_Match(image_paths):
img1 = cv2.imread(image_paths[0])
img2 = cv2.imread(image_paths[1])
# 初始化ORB检测器
orb = cv2.ORB_create()
start_time = time.time()
# 检测关键点并计算描述子
kp1, des1 = orb.detectAndCompute(img1, None)
kp2, des2 = orb.detectAndCompute(img2, None)
# 创建BFMatcher并进行匹配
bf = cv2.BFMatcher(cv2.NORM_HAMMING, crossCheck=True)
matches = bf.match(des1, des2)
# 计算绘制时间
pipei_time = time.time() - start_time
pipei_time=round(pipei_time,3)
# 按距离排序并选择前N个最佳匹配
matches_sorted = sorted(matches, key=lambda x: x.distance)
top_n_matches = matches_sorted[:draw_num] # 选择前30个最佳匹配
# 使用修改后的函数绘制匹配结果
result_img = draw_ORB_matches(img1, kp1, img2, kp2, top_n_matches)
cv2.putText(result_img, "ORB-time:"+str(pipei_time), (30, 90), cv2.FONT_HERSHEY_SIMPLEX, font_scale, color_title, thickness)
return result_img
def sift_feature_matching(image_paths, top_n=30, match_ratio=0.7):
"""
使用SIFT算法进行图像特征匹配
参数:
image_paths: 包含两个图像路径的列表
top_n: 选择最佳匹配的数量
match_ratio: 比率测试的阈值（用于FLANN匹配器）
返回:
result_img: 绘制匹配结果的图像
match_count: 成功匹配的数量
avg_distance: 平均匹配距离
"""
# 读取图像
img1 = cv2.imread(image_paths[0])
img2 = cv2.imread(image_paths[1])
if img1 is None or img2 is None:
print("错误：无法读取图像文件")
return None, 0, 0
# 转换为灰度图
gray1 = cv2.cvtColor(img1, cv2.COLOR_BGR2GRAY)
gray2 = cv2.cvtColor(img2, cv2.COLOR_BGR2GRAY)
# 创建SIFT特征检测器
sift = cv2.SIFT_create(
nfeatures=2000, # 保留的最佳特征点数量（0表示保留所有）
nOctaveLayers=3, # 每个八度空间的层数
contrastThreshold=0.04, # 对比度阈值
edgeThreshold=10, # 边缘阈值
sigma=1.6 # 高斯滤波器参数
)
start_time = time.time()
# 检测关键点并计算描述符
kp1, des1 = sift.detectAndCompute(gray1, None)
kp2, des2 = sift.detectAndCompute(gray2, None)
print(f"图像1检测到 {len(kp1)} 个关键点")
print(f"图像2检测到 {len(kp2)} 个关键点")
if des1 is None or des2 is None:
print("错误：无法计算特征描述符")
return None, 0, 0
# 创建FLANN匹配器
index_params = dict(algorithm=1, trees=5) # KD树算法
search_params = dict(checks=50) # 搜索次数
flann = cv2.FlannBasedMatcher(index_params, search_params)
# 使用kNN匹配，k=2
matches = flann.knnMatch(des1, des2, k=2)
# 计算绘制时间
pipei_time = time.time() - start_time
pipei_time=round(pipei_time,3)
# 应用比率测试筛选优质匹配
good_matches = []
for m, n in matches:
if m.distance < match_ratio * n.distance:
good_matches.append(m)
print(f"原始匹配数量: {len(matches)}")
print(f"经过比率测试后的优质匹配数量: {len(good_matches)}")
# 按匹配距离排序并选择前N个最佳匹配
good_matches_sorted = sorted(good_matches, key=lambda x: x.distance)
top_matches = good_matches_sorted[:top_n]
# 计算匹配质量指标
if len(top_matches) > 0:
avg_distance = sum(m.distance for m in top_matches) / len(top_matches)
min_distance = min(m.distance for m in top_matches)
max_distance = max(m.distance for m in top_matches)
print(f"最佳匹配距离: {min_distance:.2f}")
print(f"最差匹配距离: {max_distance:.2f}")
print(f"平均匹配距离: {avg_distance:.2f}")
else:
avg_distance = 0
print("警告：没有找到足够的优质匹配")
# 绘制匹配结果
result_img = draw_sift_matches(img1, kp1, img2, kp2, top_matches)
cv2.putText(result_img, "SIFT-time:"+str(pipei_time), (30, 90), cv2.FONT_HERSHEY_SIMPLEX, font_scale, color_title, thickness)
return result_img, len(top_matches), avg_distance
def draw_sift_matches(img1, kp1, img2, kp2, matches, line_thickness=line_thickness_set):
"""
绘制SIFT匹配结果，根据匹配质量使用渐变色
参数:
img1, img2: 输入图像
kp1, kp2: 关键点列表
matches: 匹配结果
line_thickness: 连线粗细
返回:
绘制匹配结果的图像
"""
# 获取图像尺寸
h1, w1 = img1.shape[:2]
h2, w2 = img2.shape[:2]
# 创建拼接图像
new_width = w1 + w2
new_height = max(h1, h2)
result = np.zeros((new_height, new_width, 3), dtype=np.uint8)
result[0:h1, 0:w1] = img1
result[0:h2, w1:w1+w2] = img2
if len(matches) == 0:
print("警告：匹配列表为空")
return result
# 获取所有匹配的距离值用于颜色映射
distances = [m.distance for m in matches]
min_dist = min(distances)
max_dist = max(distances)
print(f"绘制 {len(matches)} 个匹配点")
# 绘制每个匹配对
for i, match in enumerate(matches):
# 获取关键点坐标
pt1 = (int(kp1[match.queryIdx].pt[0]), int(kp1[match.queryIdx].pt[1]))
pt2 = (int(kp2[match.trainIdx].pt[0] + w1), int(kp2[match.trainIdx].pt[1]))
# # # 根据匹配质量计算颜色（距离越小，匹配质量越高）
# norm_dist = (match.distance - min_dist) / (max_dist - min_dist + 1e-5)
# # 颜色映射：高质量→绿色，中等质量→黄色，低质量→红色
# if norm_dist < 0.33: # 最好的33% - 绿色系
# blue = 0
# green = 255
# red = int(255 * (norm_dist / 0.33))
# elif norm_dist < 0.66: # 中间33% - 黄色到橙色
# blue = 0
# green = int(255 * (1 - (norm_dist - 0.33) / 0.33))
# red = 255
# else: # 最差的33% - 橙色到红色
# blue = 0
# green = int(128 * (1 - (norm_dist - 0.66) / 0.34))
# red = 255
color = colors_Line[i % len(colors_Line)]
#color = (blue, green, red)
#color=generate_pastel_color()
# 绘制连线
cv2.line(result, pt1, pt2, color, line_thickness)
# 绘制关键点
cv2.circle(result, pt1, 5, color, -1)
cv2.circle(result, pt2, 5, color, -1)
# 显示匹配序号和距离
if line_thickness > 1 and len(matches) <= 50: # 避免文本过于拥挤
cv2.putText(result, f"{i}", (pt1[0]+5, pt1[1]),
cv2.FONT_HERSHEY_SIMPLEX, 0.4, (255, 255, 255), 1)
cv2.putText(result, f"{match.distance:.1f}", (pt2[0]+5, pt2[1]),
cv2.FONT_HERSHEY_SIMPLEX, 0.4, (255, 255, 255), 1)
return result
from mast3r.model import AsymmetricMASt3R
from mast3r.fast_nn import fast_reciprocal_NNs
import mast3r.utils.path_to_dust3r
from dust3r.inference import inference
from dust3r.utils.image import load_images
import cv2
import numpy as np
import torch
import time
class FeatureMatcher_Mast3r:
def __init__(self, model_path, device='cuda'):
"""
初始化特征匹配器
Args:
model_path: MASt3R模型路径
device: 运行设备
"""
self.device = device
self.model = self.load_mast3r_model(model_path)
def load_mast3r_model(self, model_path):
"""加载MASt3R模型"""
print("开始加载MASt3R模型")
model = AsymmetricMASt3R.from_pretrained(model_path).to(self.device)
print("MASt3R模型加载完成")
return model
def load_and_preprocess_images(self, image_paths, size=512):
"""
加载并预处理图像
Args:
image_paths: 图像路径列表
size: 图像尺寸
Returns:
预处理后的图像数据
"""
images = load_images(image_paths, size=size)
# 图像标准化
image_mean = torch.as_tensor([0.5, 0.5, 0.5], device='cpu').reshape(1, 3, 1, 1)
image_std = torch.as_tensor([0.5, 0.5, 0.5], device='cpu').reshape(1, 3, 1, 1)
viz_imgs = []
for i, view in enumerate([images[0], images[1]]):
rgb_tensor = view['img'] * image_std + image_mean
viz_img = rgb_tensor.squeeze(0).permute(1, 2, 0).cpu().numpy()
viz_imgs.append(viz_img)
# 转换为OpenCV格式
img0_bgr = (viz_imgs[0][:, :, ::-1] * 255).astype(np.uint8)
img1_bgr = (viz_imgs[1][:, :, ::-1] * 255).astype(np.uint8)
return images, img0_bgr, img1_bgr
def mast3r_feature_matching(self, images):
"""
MASt3R特征匹配
Args:
images: 输入图像
Returns:
匹配结果、关键点坐标和耗时
"""
start_time = time.time()
# 执行推理
output = inference([tuple(images)], self.model, self.device, batch_size=1, verbose=False)
view1, pred1 = output['view1'], output['pred1']
view2, pred2 = output['view2'], output['pred2']
desc1, desc2 = pred1['desc'].squeeze(0).detach(), pred2['desc'].squeeze(0).detach()
# 快速互最近邻匹配
matches_im0, matches_im1 = fast_reciprocal_NNs(desc1, desc2, subsample_or_initxy1=8,
device=self.device, dist='dot', block_size=2**13)
# 有效点筛选
H0, W0 = view1['true_shape'][0]
valid_matches_im0 = (matches_im0[:, 0] >= 3) & (matches_im0[:, 0] < int(W0) - 3) & (
matches_im0[:, 1] >= 3) & (matches_im0[:, 1] < int(H0) - 3)
H1, W1 = view2['true_shape'][0]
valid_matches_im1 = (matches_im1[:, 0] >= 3) & (matches_im1[:, 0] < int(W1) - 3) & (
matches_im1[:, 1] >= 3) & (matches_im1[:, 1] < int(H1) - 3)
valid_matches = valid_matches_im0 & valid_matches_im1
matches_im0, matches_im1 = matches_im0[valid_matches], matches_im1[valid_matches]
end_time = time.time()
total_time = end_time - start_time
print("Mast3R 耗费时间：",total_time)
return matches_im0, matches_im1, total_time
def visualize_results(self, mast3r_matches_im0, mast3r_matches_im1, img0_bgr, img1_bgr, n_viz=100):
"""
可视化MASt3R匹配结果
"""
h0, w0 = img0_bgr.shape[:2]
h1, w1 = img1_bgr.shape[:2]
max_height = max(h0, h1)
# 创建结果图像
result_img = np.zeros((max_height, w0 + w1, 3), dtype=np.uint8)
# 放置原始图像
result_img[:h0, :w0] = img0_bgr
result_img[:h1, w0:w0+w1] = img1_bgr
# 绘制MASt3R匹配点
colors = [
(255, 0, 0), (0, 255, 0), (0, 0, 255), (255, 255, 0), (255, 0, 255),(0, 255, 255),(0, 100, 255),(0, 255, 100)
]
print('总点数',mast3r_matches_im0.shape[0])
n_viz_actual = min(n_viz, mast3r_matches_im0.shape[0])
indices = np.round(np.linspace(0, mast3r_matches_im0.shape[0] - 1, n_viz_actual)).astype(int)
for i, idx in enumerate(indices):
color = colors[i % len(colors)]
pt1 = mast3r_matches_im0[idx]
pt2 = mast3r_matches_im1[idx]
x1, y1 = int(pt1[0]), int(pt1[1])
x2, y2 = int(pt2[0]) + w0, int(pt2[1])
cv2.line(result_img, (x1, y1), (x2, y2), color, 1, cv2.LINE_AA)
cv2.circle(result_img, (x1, y1), 3, color, -1)
cv2.circle(result_img, (x2, y2), 3, color, -1)
# # 添加标签
# cv2.putText(result_img, "MASt3R-Time:"+str(), (30, 200),
# cv2.FONT_HERSHEY_SIMPLEX, 0.7, color, 2)
return result_img
if __name__ == "__main__":
# =============================1 执行ORB匹配
# 读取图像
image_paths = [
'/media/r9000k/DD_XS/2数据/2_1精灵4无rtk/2圆楼_滨湖_白天_resize/DJI_0516.JPG',
'/media/r9000k/DD_XS/2数据/2_1精灵4无rtk/v4_2圆楼_滨湖_拍照模式测曝光_resize/DJI_0146.JPG'
]
orb_result_img = Orb_Match(image_paths)
# 显示结果
# cv2.namedWindow('orb', cv2.WINDOW_NORMAL)
# cv2.imshow('orb', orb_result_img)
# cv2.waitKey(0)
# cv2.destroyAllWindows()
# =============================2 执行SIFT匹配
sift_result_img, match_count, avg_distance = sift_feature_matching(
image_paths,
top_n=draw_num, # 显示前30个最佳匹配
match_ratio=0.7 # 比率测试阈值
)
# 显示结果
# cv2.namedWindow('SIFT Feature Matching', cv2.WINDOW_NORMAL)
# cv2.imshow('SIFT Feature Matching', sift_result_img)
# cv2.waitKey(0)
# cv2.destroyAllWindows()
# =============================3 执行Mast3r匹配
# 创建特征匹配器实例
Mast3r_model_path = "checkpoints/MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric.pth"
matcher_Mast3r = FeatureMatcher_Mast3r(Mast3r_model_path)
# 加载图像
images, img0_bgr, img1_bgr = matcher_Mast3r.load_and_preprocess_images(image_paths)
# MASt3R特征匹配
mast3r_pts1, mast3r_pts2, mast3r_time = matcher_Mast3r.mast3r_feature_matching(images)
# 可视化结果
mast3r_result_img = matcher_Mast3r.visualize_results(mast3r_pts1, mast3r_pts2, img0_bgr, img1_bgr, n_viz=draw_num)
# 添加标签
cv2.putText(mast3r_result_img, "MASt3R-Time:"+str(round(mast3r_time,3)), (10, 30),
cv2.FONT_HERSHEY_SIMPLEX, 0.7, color_title, 2)
# 显示结果
# cv2.namedWindow('MASt3R Feature Matching Results', cv2.WINDOW_NORMAL)
# cv2.imshow('MASt3R Feature Matching Results', mast3r_result_img)
# cv2.waitKey(0)
# cv2.destroyAllWindows()
# =============================4 可视化对比图
# 如果使用 PyQt5
from PyQt5.QtCore import QLibraryInfo
import os
os.environ["QT_QPA_PLATFORM_PLUGIN_PATH"] = QLibraryInfo.location(QLibraryInfo.PluginsPath)
import matplotlib.pyplot as plt
import matplotlib.image as mpimg
# 将三张图片存入列表方便处理
images = [orb_result_img, sift_result_img, mast3r_result_img]
titles = ['ORB', 'SIFT', 'Mast3R'] # 可自定义每个子图的标题
# 2. 创建子图布局：3行1列
fig, axs = plt.subplots(3, 1, figsize=(20, 20)) # figsize可调整整个窗口大小
# 3. 在每个子图中显示图片
for i, ax in enumerate(axs):
result_img_rgb = images[i][:, :, ::-1] # BGR转RGB OpenCV 的 BGR 格式与 matplotlib 的 RGB 格式
ax.imshow(result_img_rgb) # 显示图片
ax.set_title(titles[i]) # 设置标题
ax.axis('off') # 关闭坐标轴，让图片更清爽
# 4. 调整布局并显示
plt.tight_layout()
plt.show()
浙公网安备 33010602011771号