PaddlePaddle-Detection-YOLOv3的NMS算法

先定义计算IOU的函数
注意：PaddleDetection::ObjectResult的输出tensor为[num_id, score, xmin, ymin, xmax, ymax]

//定义计算iou的函数
float get_iou_value(PaddleDetection::ObjectResult box2_info, PaddleDetection::ObjectResult box1_info) {
	float iou_value;
	//计算iou_value，先寻找两个框左上角的最大值，再寻找两个框右下角的最小值
	float x1 = max(box2_info.rect[0], box1_info.rect[0]);
	float y1 = max(box2_info.rect[1], box1_info.rect[1]);
	float x2 = min(box2_info.rect[2], box1_info.rect[2]);
	float y2 = min(box2_info.rect[3], box1_info.rect[3]);

	float w = max(x2 - x1, 0);
	float h = max(y2 - y1, 0);
	float insection_area = w * h;

	float box1_area = (box1_info.rect[2] - box1_info.rect[0])*(box1_info.rect[3] - box1_info.rect[1]);
	float box2_area = (box2_info.rect[2] - box2_info.rect[0])*(box2_info.rect[3] - box2_info.rect[1]);
	float union_area = box1_area + box2_area - insection_area;
	iou_value = insection_area / union_area;

	return iou_value;
}

NMS算法实现：

//1.根据置信度，利用冒泡排序算法对检测结果进行排序
int boxes_num = im_result.size();
PaddleDetection::ObjectResult item_;
for (int i = 0; i < boxes_num -1; i++) {
	for (int j = 0; j < boxes_num - 1 - i; j++) {
		if (im_result[j].confidence < im_result[j + 1].confidence) {
			item_ = im_result[j];
			im_result[j] = im_result[j+1];
			im_result[j + 1] = item_;
		}
	}
}


//2.依次计算相邻两个box的iou，若iou大于设定阈值，则过滤低置信度的box
float conf_threshold = 0.7;

for (int i = 0; i < boxes_num - 1; ++i) {
	for (int j = i+1; j < boxes_num; ++j) {
		//计算iou
		float iou_value = get_iou_value(im_result[i], im_result[j]);
		if (iou_value > conf_threshold) {
			im_result.erase(im_result.begin() + j);
			--j;
			--boxes_num;
		}
	}
}