ncnn/examples/scrfd_crowdhuman.cpp

// Copyright 2021 Tencent
// SPDX-License-Identifier: BSD-3-Clause

#include "net.h"

#if defined(USE_NCNN_SIMPLEOCV)
#include "simpleocv.h"
#else
#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>
#endif
#include <stdio.h>
#include <vector>

struct FaceObject
{
    cv::Rect_<float> rect;
    float prob;
};

static inline float intersection_area(const FaceObject& a, const FaceObject& b)
{
    cv::Rect_<float> inter = a.rect & b.rect;
    return inter.area();
}

static void qsort_descent_inplace(std::vector<FaceObject>& faceobjects, int left, int right)
{
    int i = left;
    int j = right;
    float p = faceobjects[(left + right) / 2].prob;

    while (i <= j)
    {
        while (faceobjects[i].prob > p)
            i++;

        while (faceobjects[j].prob < p)
            j--;

        if (i <= j)
        {
            // swap
            std::swap(faceobjects[i], faceobjects[j]);

            i++;
            j--;
        }
    }

    #pragma omp parallel sections
    {
        #pragma omp section
        {
            if (left < j) qsort_descent_inplace(faceobjects, left, j);
        }
        #pragma omp section
        {
            if (i < right) qsort_descent_inplace(faceobjects, i, right);
        }
    }
}

static void qsort_descent_inplace(std::vector<FaceObject>& faceobjects)
{
    if (faceobjects.empty())
        return;

    qsort_descent_inplace(faceobjects, 0, faceobjects.size() - 1);
}

static void nms_sorted_bboxes(const std::vector<FaceObject>& faceobjects, std::vector<int>& picked, float nms_threshold)
{
    picked.clear();

    const int n = faceobjects.size();

    std::vector<float> areas(n);
    for (int i = 0; i < n; i++)
    {
        areas[i] = faceobjects[i].rect.area();
    }

    for (int i = 0; i < n; i++)
    {
        const FaceObject& a = faceobjects[i];

        int keep = 1;
        for (int j = 0; j < (int)picked.size(); j++)
        {
            const FaceObject& b = faceobjects[picked[j]];

            // intersection over union
            float inter_area = intersection_area(a, b);
            float union_area = areas[i] + areas[picked[j]] - inter_area;
            //             float IoU = inter_area / union_area
            if (inter_area / union_area > nms_threshold)
                keep = 0;
        }

        if (keep)
            picked.push_back(i);
    }
}

// insightface/detection/scrfd/mmdet/core/anchor/anchor_generator.py gen_single_level_base_anchors()
static ncnn::Mat generate_anchors(int base_size, const ncnn::Mat& ratios, const ncnn::Mat& scales)
{
    int num_ratio = ratios.w;
    int num_scale = scales.w;

    ncnn::Mat anchors;
    anchors.create(4, num_ratio * num_scale);

    const float cx = 0;
    const float cy = 0;

    for (int i = 0; i < num_ratio; i++)
    {
        float ar = ratios[i];

        int r_w = round(base_size / sqrt(ar));
        int r_h = round(r_w * ar); //round(base_size * sqrt(ar));

        for (int j = 0; j < num_scale; j++)
        {
            float scale = scales[j];

            float rs_w = r_w * scale;
            float rs_h = r_h * scale;

            float* anchor = anchors.row(i * num_scale + j);

            anchor[0] = cx - rs_w * 0.5f;
            anchor[1] = cy - rs_h * 0.5f;
            anchor[2] = cx + rs_w * 0.5f;
            anchor[3] = cy + rs_h * 0.5f;
        }
    }

    return anchors;
}

static void generate_proposals(const ncnn::Mat& anchors, int feat_stride, const ncnn::Mat& score_blob, const ncnn::Mat& bbox_blob, float prob_threshold, std::vector<FaceObject>& faceobjects)
{
    int w = score_blob.w;
    int h = score_blob.h;

    // generate face proposal from bbox deltas and shifted anchors
    const int num_anchors = anchors.h;

    for (int q = 0; q < num_anchors; q++)
    {
        const float* anchor = anchors.row(q);

        const ncnn::Mat score = score_blob.channel(q);
        const ncnn::Mat bbox = bbox_blob.channel_range(q * 4, 4);

        // shifted anchor
        float anchor_y = anchor[1];

        float anchor_w = anchor[2] - anchor[0];
        float anchor_h = anchor[3] - anchor[1];

        for (int i = 0; i < h; i++)
        {
            float anchor_x = anchor[0];

            for (int j = 0; j < w; j++)
            {
                int index = i * w + j;

                float prob = score[index];

                if (prob >= prob_threshold)
                {
                    // insightface/detection/scrfd/mmdet/models/dense_heads/scrfd_head.py _get_bboxes_single()
                    float dx = bbox.channel(0)[index] * feat_stride;
                    float dy = bbox.channel(1)[index] * feat_stride;
                    float dw = bbox.channel(2)[index] * feat_stride;
                    float dh = bbox.channel(3)[index] * feat_stride;

                    // insightface/detection/scrfd/mmdet/core/bbox/transforms.py distance2bbox()
                    float cx = anchor_x + anchor_w * 0.5f;
                    float cy = anchor_y + anchor_h * 0.5f;

                    float x0 = cx - dx;
                    float y0 = cy - dy;
                    float x1 = cx + dw;
                    float y1 = cy + dh;

                    FaceObject obj;
                    obj.rect.x = x0;
                    obj.rect.y = y0;
                    obj.rect.width = x1 - x0 + 1;
                    obj.rect.height = y1 - y0 + 1;
                    obj.prob = prob;

                    faceobjects.push_back(obj);
                }

                anchor_x += feat_stride;
            }

            anchor_y += feat_stride;
        }
    }
}

static int detect_scrfd(const cv::Mat& bgr, std::vector<FaceObject>& faceobjects)
{
    ncnn::Net scrfd;

    scrfd.opt.use_vulkan_compute = true;

    // Insight face does not provided a trained scrfd_crowdhuman model
    // but I have one for detecing cat face, you can have a try here:
    // https://drive.google.com/file/d/1JogkKa0f_09HkENbCnXy9hRYxm35wKTn

    if (scrfd.load_param("scrfd_crowdhuman.param"))
        exit(-1);
    if (scrfd.load_model("scrfd_crowdhuman.bin"))
        exit(-1);

    int width = bgr.cols;
    int height = bgr.rows;

    const int target_size = 640;
    const float prob_threshold = 0.3f;
    const float nms_threshold = 0.45f;

    // pad to multiple of 32
    int w = width;
    int h = height;
    float scale = 1.f;
    if (w > h)
    {
        scale = (float)target_size / w;
        w = target_size;
        h = h * scale;
    }
    else
    {
        scale = (float)target_size / h;
        h = target_size;
        w = w * scale;
    }

    ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, width, height, w, h);

    // pad to target_size rectangle
    int wpad = (w + 31) / 32 * 32 - w;
    int hpad = (h + 31) / 32 * 32 - h;
    ncnn::Mat in_pad;
    ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 0.f);

    const float mean_vals[3] = {127.5f, 127.5f, 127.5f};
    const float norm_vals[3] = {1 / 128.f, 1 / 128.f, 1 / 128.f};
    in_pad.substract_mean_normalize(mean_vals, norm_vals);

    ncnn::Extractor ex = scrfd.create_extractor();

    ex.input("input.1", in_pad);

    std::vector<FaceObject> faceproposals;

    // stride 8
    {
        ncnn::Mat score_blob, bbox_blob;
        ex.extract("490", score_blob);
        ex.extract("493", bbox_blob);

        const int base_size = 8;
        const int feat_stride = 8;
        ncnn::Mat ratios(1);
        ratios[0] = 2.f;
        ncnn::Mat scales(1);
        scales[0] = 3.f;
        ncnn::Mat anchors = generate_anchors(base_size, ratios, scales);

        std::vector<FaceObject> faceobjects32;
        generate_proposals(anchors, feat_stride, score_blob, bbox_blob, prob_threshold, faceobjects32);

        faceproposals.insert(faceproposals.end(), faceobjects32.begin(), faceobjects32.end());
    }

    // stride 16
    {
        ncnn::Mat score_blob, bbox_blob;
        ex.extract("510", score_blob);
        ex.extract("513", bbox_blob);

        const int base_size = 16;
        const int feat_stride = 16;
        ncnn::Mat ratios(1);
        ratios[0] = 2.f;
        ncnn::Mat scales(1);
        scales[0] = 3.f;
        ncnn::Mat anchors = generate_anchors(base_size, ratios, scales);

        std::vector<FaceObject> faceobjects16;
        generate_proposals(anchors, feat_stride, score_blob, bbox_blob, prob_threshold, faceobjects16);

        faceproposals.insert(faceproposals.end(), faceobjects16.begin(), faceobjects16.end());
    }

    // stride 32
    {
        ncnn::Mat score_blob, bbox_blob;
        ex.extract("530", score_blob);
        ex.extract("533", bbox_blob);

        const int base_size = 32;
        const int feat_stride = 32;
        ncnn::Mat ratios(1);
        ratios[0] = 2.f;
        ncnn::Mat scales(1);
        scales[0] = 3.f;
        ncnn::Mat anchors = generate_anchors(base_size, ratios, scales);

        std::vector<FaceObject> faceobjects8;
        generate_proposals(anchors, feat_stride, score_blob, bbox_blob, prob_threshold, faceobjects8);

        faceproposals.insert(faceproposals.end(), faceobjects8.begin(), faceobjects8.end());
    }

    // stride 64
    {
        ncnn::Mat score_blob, bbox_blob, kps_blob;
        ex.extract("550", score_blob);
        ex.extract("553", bbox_blob);

        const int base_size = 64;
        const int feat_stride = 64;
        ncnn::Mat ratios(1);
        ratios[0] = 2.f;
        ncnn::Mat scales(1);
        scales[0] = 3.f;
        ncnn::Mat anchors = generate_anchors(base_size, ratios, scales);

        std::vector<FaceObject> faceobjects8;
        generate_proposals(anchors, feat_stride, score_blob, bbox_blob, prob_threshold, faceobjects8);

        faceproposals.insert(faceproposals.end(), faceobjects8.begin(), faceobjects8.end());
    }

    // stride 128
    {
        ncnn::Mat score_blob, bbox_blob, kps_blob;
        ex.extract("570", score_blob);
        ex.extract("573", bbox_blob);

        const int base_size = 128;
        const int feat_stride = 128;
        ncnn::Mat ratios(1);
        ratios[0] = 2.f;
        ncnn::Mat scales(1);
        scales[0] = 3.f;
        ncnn::Mat anchors = generate_anchors(base_size, ratios, scales);

        std::vector<FaceObject> faceobjects8;
        generate_proposals(anchors, feat_stride, score_blob, bbox_blob, prob_threshold, faceobjects8);

        faceproposals.insert(faceproposals.end(), faceobjects8.begin(), faceobjects8.end());
    }

    // sort all proposals by score from highest to lowest
    qsort_descent_inplace(faceproposals);

    // apply nms with nms_threshold
    std::vector<int> picked;
    nms_sorted_bboxes(faceproposals, picked, nms_threshold);

    int face_count = picked.size();

    faceobjects.resize(face_count);
    for (int i = 0; i < face_count; i++)
    {
        faceobjects[i] = faceproposals[picked[i]];

        // adjust offset to original unpadded
        float x0 = (faceobjects[i].rect.x - (wpad / 2)) / scale;
        float y0 = (faceobjects[i].rect.y - (hpad / 2)) / scale;
        float x1 = (faceobjects[i].rect.x + faceobjects[i].rect.width - (wpad / 2)) / scale;
        float y1 = (faceobjects[i].rect.y + faceobjects[i].rect.height - (hpad / 2)) / scale;

        x0 = std::max(std::min(x0, (float)width - 1), 0.f);
        y0 = std::max(std::min(y0, (float)height - 1), 0.f);
        x1 = std::max(std::min(x1, (float)width - 1), 0.f);
        y1 = std::max(std::min(y1, (float)height - 1), 0.f);

        faceobjects[i].rect.x = x0;
        faceobjects[i].rect.y = y0;
        faceobjects[i].rect.width = x1 - x0;
        faceobjects[i].rect.height = y1 - y0;
    }

    return 0;
}

static void draw_faceobjects(const cv::Mat& bgr, const std::vector<FaceObject>& faceobjects)
{
    cv::Mat image = bgr.clone();

    for (size_t i = 0; i < faceobjects.size(); i++)
    {
        const FaceObject& obj = faceobjects[i];

        fprintf(stderr, "%.5f at %.2f %.2f %.2f x %.2f\n", obj.prob,
                obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);

        cv::rectangle(image, obj.rect, cv::Scalar(0, 255, 0));

        char text[256];
        sprintf(text, "%.1f%%", obj.prob * 100);

        int baseLine = 0;
        cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);

        int x = obj.rect.x;
        int y = obj.rect.y - label_size.height - baseLine;
        if (y < 0)
            y = 0;
        if (x + label_size.width > image.cols)
            x = image.cols - label_size.width;

        cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),
                      cv::Scalar(255, 255, 255), -1);

        cv::putText(image, text, cv::Point(x, y + label_size.height),
                    cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));
    }

    cv::imshow("image", image);
    cv::waitKey(0);
}

int main(int argc, char** argv)
{
    if (argc != 2)
    {
        fprintf(stderr, "Usage: %s [imagepath]\n", argv[0]);
        return -1;
    }

    const char* imagepath = argv[1];

    cv::Mat m = cv::imread(imagepath, 1);
    if (m.empty())
    {
        fprintf(stderr, "cv::imread %s failed\n", imagepath);
        return -1;
    }

    std::vector<FaceObject> faceobjects;
    detect_scrfd(m, faceobjects);

    draw_faceobjects(m, faceobjects);

    return 0;
}