Error with 'out1.to_pixels(score.data, ncnn::Mat::PIXEL_GRAY)'

[7Ston1]

Hello, Linyi. Thank you very much for this outstanding work. However, I recently discovered a issue during testing. The code 'out1.to_pixels(score.data, ncnn::Mat::PIXEL_GRAY);' doesn't seem to work while the format 'ncnn::Mat::PIXEL_BGR' can work well. I noticed that you tried to use 'memcpy()' to solve the problem. Do you know why cause this?

[ffff349]

我也发现的这个问题，导致我的特征点出不来

[deepConnectionism]

@7Ston1 这个问题是因为 输入的是 bgr 三通道图像，你需要修改：

void FeatureTracker::let_net(const cv::Mat& image_bgr) {
    last_desc = desc.clone();

    // 将 BGR 图像转换为灰度图像
    cv::Mat img_gray;
    cv::cvtColor(image_bgr, img_gray, cv::COLOR_BGR2GRAY);

    // 调整灰度图像大小
    cv::Mat img_resized;
    cv::resize(img_gray, img_resized, cv::Size(LET_WIDTH, LET_HEIGHT));

    // 创建 NCNN 提取器
    ncnn::Extractor ex = net.create_extractor();
    ex.set_light_mode(true);

     // 将灰度图像转换为 NCNN Mat 格式
    in = ncnn::Mat::from_pixels(img_resized.data, ncnn::Mat::PIXEL_GRAY, img_resized.cols, img_resized.rows);
    in.substract_mean_normalize(mean_vals_gray, norm_vals_gray);

    // 提取特征
    ex.input("input", in);
    ex.extract("score", out1);
    ex.extract("descriptor", out2);

    // 逆归一化处理
    out1.substract_mean_normalize(mean_vals_inv_gray, norm_vals_inv_gray);
    out2.substract_mean_normalize(mean_vals_inv_gray, norm_vals_inv_gray);

    // 将 NCNN Mat 转换为 OpenCV Mat
    memcpy((uchar*)score.data, out1.data, LET_HEIGHT*LET_WIDTH*sizeof(float));
    cv::Mat desc_tmp(LET_HEIGHT, LET_WIDTH, CV_8UC3);
    out2.to_pixels(desc_tmp.data, ncnn::Mat::PIXEL_BGR);
    desc = desc_tmp.clone();
    
    // // 保存描述子图像
    // cv::imwrite("desc.png", desc);
}

此外，你需要使用 灰度模型