如何在C++(OpenCV)中堆叠2D Mat生成可遍历的3D数组？

Great question! Coming from NumPy's handy vstack/stack functions, it’s totally reasonable to want a similar workflow in OpenCV for C++. Luckily, OpenCV’s cv::Mat natively supports multi-dimensional arrays, so we can achieve exactly what you need—stacking a 2D grayscale image into a 3D matrix with N copies, and easily traversing each element to modify values.

Let’s walk through two straightforward approaches:

1. Quick & Clean: Use cv::repeat + cv::reshape #

This method is perfect if you just need to duplicate your image N times and reshape the result into a 3D array in one go. It’s concise and requires minimal code.

Example Code #

#include <opencv2/opencv.hpp>
#include <iostream>

int main() {
    // Load your grayscale image
    cv::Mat im = cv::imread("image.png", 0);
    if (im.empty()) {
        std::cerr << "Failed to load image!" << std::endl;
        return -1;
    }

    const int N = 5; // Number of layers you want to stack

    // Step 1: Repeat the image N times vertically (result is H*N x W, single channel)
    cv::Mat repeated = cv::repeat(im, N, 1);

    // Step 2: Reshape into a 3D Mat with dimensions (N, image_height, image_width)
    // reshape params: 1st = number of channels (keep 1), 2nd = new 3D size vector
    cv::Mat stacked_3d = repeated.reshape(1, cv::Size({N, im.rows, im.cols}));

    // Now you have your 3D matrix ready for use!
    return 0;
}

2. Manual Control: Create 3D Mat & Fill Layers #

If you need more flexibility (e.g., modifying layers as you stack them), manually creating the 3D matrix and copying each layer gives you full control over the process.

Example Code #

#include <opencv2/opencv.hpp>
#include <iostream>

int main() {
    cv::Mat im = cv::imread("image.png", 0);
    if (im.empty()) {
        std::cerr << "Failed to load image!" << std::endl;
        return -1;
    }

    const int N = 5;
    // Define 3D dimensions: (number of layers, image height, image width)
    int dims[] = {N, im.rows, im.cols};

    // Create an empty 3D Mat with the same data type as your input image
    cv::Mat stacked_3d(3, dims, im.type());

    // Copy the original image into each layer
    for (int layer_idx = 0; layer_idx < N; ++layer_idx) {
        // Extract the current layer as a sub-matrix
        cv::Mat layer = stacked_3d(cv::Range(layer_idx, layer_idx + 1), 
                                   cv::Range::all(), 
                                   cv::Range::all());
        // Reshape the layer to match the original image's 2D size, then copy
        im.copyTo(layer.reshape(1, im.rows));
    }

    return 0;
}

Traversing the 3D Mat #

Once you have your 3D matrix, accessing and modifying elements is straightforward using cv::Mat::at<>(). For a grayscale image (uchar type), you can iterate through all layers, rows, and columns like this:

// Iterate through every element in the 3D matrix
for (int i = 0; i < stacked_3d.size[0]; ++i) { // i = layer index
    for (int y = 0; y < stacked_3d.size[1]; ++y) { // y = row
        for (int x = 0; x < stacked_3d.size[2]; ++x) { // x = column
            // Add 1 to each element (example operation)
            stacked_3d.at<uchar>(i, y, x) += 1;
        }
    }
}

Quick Notes #

• If you prefer a multi-channel 2D Mat instead of a strict 3D array (e.g., treating layers as channels), you could use cv::merge with a vector of repeated images—but this is a different structure than a true 3D matrix.
• For better performance with large images, consider using pointer-based access instead of at<>(), but at<>() is perfectly fine for most use cases and much more readable.

内容的提问来源于stack exchange，提问作者daniglezad