问题描述

目标检测模型输出的检测框存在内嵌情况。

---

原因分析与解决方法：

根据经验，第一感觉是后处理nms部分出了问题。来看下对应的代码：

static float CalcIou(const vector<float> &box1, const vector<float> &box2)
{
    float area1 = box1[6];
    float area2 = box2[6];
    float xx1 = max(box1[0], box2[0]);
    float yy1 = max(box1[1], box2[1]);
    float xx2 = min(box1[2], box2[2]);
    float yy2 = min(box1[3], box2[3]);
    float w = max(0.0f, xx2 - xx1 + 1);
    float h = max(0.0f, yy2 - yy1 + 1);
    float inter = w * h;
    float ovr = inter /(area1 + area2 - inter);
    return ovr;
}

static void MulticlassNms(vector<vector<float>>& bboxes, const vector<vector<float>>& vaildBox, float nmsThr)
{
    for (auto &item : vaildBox) { /* score, xcenter, ycenter, w, h, classId */
        float boxXCenter = item[XCENTER_IDX];
        float boxYCenter = item[YCENTER_IDX];
        float boxWidth = item[W_IDX];
        float boxHeight = item[H_IDX];

        float x1 = (boxXCenter - boxWidth / 2);
        float y1 = (boxYCenter - boxHeight / 2);
        float x2 = (boxXCenter + boxWidth / 2);
        float y2 = (boxYCenter + boxHeight / 2);
        float area = (x2 - x1 + 1) * (y2 - y1 + 1);
        bool keep = true;
        /* lx, ly, rx, ry, score, class id, area */
        vector<float> bbox {x1, y1, x2, y2, item[SCORE_IDX], item[CLSAA_ID_IDX], area};
        for (size_t j = 0; j < bboxes.size(); j++) {
            if (CalcIou(bbox, bboxes[j]) > nmsThr) {
                keep = false;
                break;
            }
        }
        if (keep) {
            bboxes.push_back(bbox);
        }
    }
}

目前分析最可能的原因是nms的nmsThr设置过大，没能滤除重叠检测框，原来nmsThr设置的为0.45，现调整为0.1。
检测框内嵌情况基本消失：

---

后续及思考

先给个结论，综合的看下各个Loss函数的不同点：：
IOU_Loss：主要考虑检测框和目标框重叠面积。
GIOU_Loss：在IOU的基础上，解决边界框不重合时的问题。
DIOU_Loss：在IOU和GIOU的基础上，考虑边界框中心点距离的信息。
CIOU_Loss：在DIOU的基础上，考虑边界框宽高比的尺度信息。
此项目中用的是基本的IOU，在推理性能足够的情况下，可以考虑使用DIOU,下面也给出使用DIOU的nms代码：

static float CalcDiou(const vector<float>& box1, const vector<float>& box2) {
    float x1 = min(box1[0], box2[0]);
    float y1 = min(box1[1], box2[1]);
    float x2 = max(box1[2], box2[2]);
    float y2 = max(box1[3], box2[3]);
    
    float c_x1 = (box1[0] + box1[2]) / 2.0;
    float c_y1 = (box1[1] + box1[3]) / 2.0;
    float c_x2 = (box2[0] + box2[2]) / 2.0;
    float c_y2 = (box2[1] + box2[3]) / 2.0;
    
    float dist_center = sqrt((c_x1 - c_x2) * (c_x1 - c_x2) + (c_y1 - c_y2) * (c_y1 - c_y2));
    
    float w = max(0.0f, x2 - x1);
    float h = max(0.0f, y2 - y1);
    
    float intersection = w * h;
    float area1 = (box1[2] - box1[0]) * (box1[3] - box1[1]);
    float area2 = (box2[2] - box2[0]) * (box2[3] - box2[1]);
    
    float union_area = area1 + area2 - intersection;
    
    float diou = intersection / union_area - dist_center * dist_center / (union_area * union_area);
    
    return diou;
}

static void MulticlassNms(vector<vector<float>>& bboxes, const vector<vector<float>>& vaildBox, float nmsThr)
{
    for (auto &item : vaildBox) { /* score, xcenter, ycenter, w, h, classId */
        float boxXCenter = item[XCENTER_IDX];
        float boxYCenter = item[YCENTER_IDX];
        float boxWidth = item[W_IDX];
        float boxHeight = item[H_IDX];

        float x1 = (boxXCenter - boxWidth / 2);
        float y1 = (boxYCenter - boxHeight / 2);
        float x2 = (boxXCenter + boxWidth / 2);
        float y2 = (boxYCenter + boxHeight / 2);
        float area = (x2 - x1 + 1) * (y2 - y1 + 1);
        bool keep = true;

        vector<float> bbox {x1, y1, x2, y2, item[SCORE_IDX], item[CLSAA_ID_IDX], area};
        for (size_t j = 0; j < bboxes.size(); j++) {
            if (CalcDiou(bbox, bboxes[j]) > nmsThr) {
                keep = false;
                break;
            }
        }
        if (keep) {
            bboxes.push_back(bbox);
        }
    }
}

有读者会有疑问，这里为什么不用CIOU_nms，而用DIOU_nms?
答：因为CIOU_loss，是在DIOU_loss的基础上，添加的影响因子，包含groundtruth标注框的信息，在训练时用于回归。
但在测试过程中，并没有groundtruth的信息，不用考虑影响因子，因此直接用DIOU_nms即可。

参考文档

https://blog.csdn.net/nan355655600/article/details/106246625