OpenVINO部署Mask-RCNN实例分割网络

OpenCV学堂

发布于 2021-03-12 16:33:06

7410

发布于 2021-03-12 16:33:06

文章被收录于专栏：贾志刚-OpenCV学堂贾志刚-OpenCV学堂

模型介绍

OpenVINO支持Mask-RCNN与yolact两种实例分割模型的部署，其中Mask-RCNN系列的实例分割网络是OpenVINO官方自带的，直接下载即可，yolact是来自第三方的公开模型库。

这里以instance-segmentation-security-0050模型为例说明，该模型基于COCO数据集训练，支持80个类别的实例分割，加上背景为81个类别。

OpenVINO支持部署Faster-RCNN与Mask-RCNN网络时候输入的解析都是基于两个输入层，它们分别是：

im_data : NCHW=[1x3x480x480]
im_info: 1x3 三个值分别是H、W、Scale=1.0

输出有四个，名称与输出格式及解释如下：

name: classes, shape: [100, ] 预测的100个类别可能性，值在[0~1]之间
name: scores: shape: [100, ] 预测的100个Box可能性，值在[0~1]之间
name: boxes, shape: [100, 4] 预测的100个Box坐标，左上角与右下角，基于输入的480x480
name: raw_masks, shape: [100, 81, 28, 28] Box ROI区域的实例分割输出，81表示类别（包含背景），28x28表示ROI大小。

上面都是官方文档给我的关于模型的相关信息，但是我发现该模型的实际推理输raw_masks输出格式大小为：100x81x14x14，这个算文档没更新吗？

代码演示

这边的代码输出层跟输入层都不止一个，所以为了简化，我用了两个for循环设置了输入与输出数据精度，然后直接通过hardcode来获取推理之后各个输出层对应的数据部分，首先获取类别，根据类别ID与Box的索引，直接获取实例分割mask，然后随机生成颜色，基于mask实现与原图BOX ROI的叠加，产生了实例分割之后的效果输出。完整的演示代码分为下面几步：

IE引擎初始化与模型加载

InferenceEngine::Core ie;
std::vector<std::string> coco_labels;
read_coco_labels(coco_labels);
cv::RNG rng(12345);

cv::Mat src = cv::imread("D:/images/sport-girls.png");
cv::namedWindow("input", cv::WINDOW_AUTOSIZE);
int im_h = src.rows;
int im_w = src.cols;

InferenceEngine::CNNNetwork network = ie.ReadNetwork(xml, bin);
InferenceEngine::InputsDataMap inputs = network.getInputsInfo();
InferenceEngine::OutputsDataMap outputs = network.getOutputsInfo();

设置输入与输出数据格式

std::string image_input_name = "";
std::string image_info_name = "";
int in_index = 0;
for (auto item : inputs) {
    if (in_index == 0) {
        image_input_name = item.first;
        auto input_data = item.second;
        input_data->setPrecision(Precision::U8);
        input_data->setLayout(Layout::NCHW);
    }
    else {
        image_info_name = item.first;
        auto input_data = item.second;
        input_data->setPrecision(Precision::FP32);
    }
    in_index++;
}

for (auto item : outputs) {
    std::string output_name = item.first;
    auto output_data = item.second;
    output_data->setPrecision(Precision::FP32);
    std::cout << "output name: " << output_name << std::endl;
}

设置blob输入数据与推理

auto executable_network = ie.LoadNetwork(network, "CPU");
auto infer_request = executable_network.CreateInferRequest();

auto input = infer_request.GetBlob(image_input_name);
matU8ToBlob<uchar>(src, input);

auto input2 = infer_request.GetBlob(image_info_name);
auto imInfoDim = inputs.find(image_info_name)->second->getTensorDesc().getDims()[1];
InferenceEngine::MemoryBlob::Ptr minput2 = InferenceEngine::as<InferenceEngine::MemoryBlob>(input2);
auto minput2Holder = minput2->wmap();
float *p = minput2Holder.as<InferenceEngine::PrecisionTrait<InferenceEngine::Precision::FP32>::value_type *>();
p[0] = static_cast<float>(inputs[image_input_name]->getTensorDesc().getDims()[2]);
p[1] = static_cast<float>(inputs[image_input_name]->getTensorDesc().getDims()[3]);
p[2] = 1.0f;

infer_request.Infer();

解析输出结果

auto scores = infer_request.GetBlob("scores");
auto boxes = infer_request.GetBlob("boxes");
auto clazzes = infer_request.GetBlob("classes");
auto raw_masks = infer_request.GetBlob("raw_masks");
const float* score_data = static_cast<PrecisionTrait<Precision::FP32>::value_type*>(scores->buffer());
const float* boxes_data = static_cast<PrecisionTrait<Precision::FP32>::value_type*>(boxes->buffer());
const float* clazzes_data = static_cast<PrecisionTrait<Precision::FP32>::value_type*>(clazzes->buffer());
const auto raw_masks_data = static_cast<PrecisionTrait<Precision::FP32>::value_type*>(raw_masks->buffer());
const SizeVector scores_outputDims = scores->getTensorDesc().getDims();
const SizeVector boxes_outputDims = boxes->getTensorDesc().getDims();
const SizeVector mask_outputDims = raw_masks->getTensorDesc().getDims();
const int max_count = scores_outputDims[0];
const int object_size = boxes_outputDims[1];
printf("mask NCHW=[%d, %d, %d, %d]\n", mask_outputDims[0], mask_outputDims[1], mask_outputDims[2], mask_outputDims[3]);
int mask_h = mask_outputDims[2];
int mask_w = mask_outputDims[3];
size_t box_stride = mask_h * mask_w * mask_outputDims[1];
for (int n = 0; n < max_count; n++) {
    float confidence = score_data[n];
    float xmin = boxes_data[n*object_size] * w_rate;
    float ymin = boxes_data[n*object_size + 1] * h_rate;
    float xmax = boxes_data[n*object_size + 2] * w_rate;
    float ymax = boxes_data[n*object_size + 3] * h_rate;
    if (confidence > 0.5) {
        cv::Scalar color(rng.uniform(0, 255), rng.uniform(0, 255), rng.uniform(0, 255));
        cv::Rect box;
        float x1 = std::min(std::max(0.0f, xmin), static_cast<float>(im_w));
        float y1 = std::min(std::max(0.0f,ymin), static_cast<float>(im_h));
        float x2 = std::min(std::max(0.0f, xmax), static_cast<float>(im_w));
        float y2 = std::min(std::max(0.0f, ymax), static_cast<float>(im_h));
        box.x = static_cast<int>(x1);
        box.y = static_cast<int>(y1);
        box.width = static_cast<int>(x2 - x1);
        box.height = static_cast<int>(y2 - y1);
        int label = static_cast<int>(clazzes_data[n]);
        std::cout <<"confidence: "<< confidence<<" class name: "<< coco_labels[label] << std::endl;
        // 解析mask
        float* mask_arr = raw_masks_data + box_stride * n + mask_h * mask_w * label;
        cv::Mat mask_mat(mask_h, mask_w, CV_32FC1, mask_arr);
        cv::Mat roi_img = src(box);
        cv::Mat resized_mask_mat(box.height, box.width, CV_32FC1);
        cv::resize(mask_mat, resized_mask_mat, cv::Size(box.width, box.height));
        cv::Mat uchar_resized_mask(box.height, box.width, CV_8UC3,color);
        roi_img.copyTo(uchar_resized_mask, resized_mask_mat <= 0.5);
        cv::addWeighted(uchar_resized_mask, 0.7, roi_img, 0.3, 0.0f, roi_img);
        cv::putText(src, coco_labels[label].c_str(), box.tl()+(box.br()-box.tl())/2, cv::FONT_HERSHEY_PLAIN, 1.0, cv::Scalar(0, 0, 255), 1, 8);
    }
}

最终程序测试结果：