Method postprocess

efficientdet/efficientdet.cpp:168–200

}

void EfficientDet::postprocess(const std::vector<cv::Mat>& imgsBatch)
{
    CHECK(cudaMemcpy(m_output_num_host,     m_output_num_device,     sizeof(int) * m_param.batch_size, cudaMemcpyDeviceToHost));
    CHECK(cudaMemcpy(m_output_boxes_host,   m_output_boxes_device,   sizeof(int) * m_param.batch_size * 1 * m_param.topK * 4, cudaMemcpyDeviceToHost));
    const auto boxes = reinterpret_cast<const float*>(m_output_boxes_host);
    CHECK(cudaMemcpy(m_output_scores_host,  m_output_scores_device,  sizeof(int) * m_param.batch_size * 1 * m_param.topK, cudaMemcpyDeviceToHost));
    const auto scores = reinterpret_cast<const float*>(m_output_scores_host);
    CHECK(cudaMemcpy(m_output_classes_host, m_output_classes_device, sizeof(int) * m_param.batch_size * 1 * m_param.topK, cudaMemcpyDeviceToHost));
    for (int bi = 0; bi < imgsBatch.size(); bi++)
    {
        for (int i = 0; i < m_output_num_host[bi]; i++)
        {
            float y1 = boxes[0 + i * 4 + bi * m_param.topK * 4];
            float x1 = boxes[1 + i * 4 + bi * m_param.topK * 4];
            float y2 = boxes[2 + i * 4 + bi * m_param.topK * 4];
            float x2 = boxes[3 + i * 4 + bi * m_param.topK * 4];

            float y_lt = m_dst2src.v3 * x1 + m_dst2src.v4 * y1 + m_dst2src.v5;
            float x_lt = m_dst2src.v0 * x1 + m_dst2src.v1 * y1 + m_dst2src.v2;
            float y_rb = m_dst2src.v3 * x2 + m_dst2src.v4 * y2 + m_dst2src.v5;
            float x_rb = m_dst2src.v0 * x2 + m_dst2src.v1 * y2 + m_dst2src.v2;

            float score = scores[i + bi * m_param.topK];
            if (score < m_param.conf_thresh)
            {
                continue;
            }
            int32_t class_id = m_output_classes_host[i + bi * m_param.topK];
            assert(class_id >= 0);  
            m_objectss[bi].emplace_back(x_lt, y_lt, x_rb, y_rb, score, class_id);
        }
    }
}

std::vector<std::vector<utils::Box>> EfficientDet::getObjectss() const
{