此外，计算bbox误差时，不是比较四个角的坐标，而是tx，ty，tw，th，具体计算如下：
预测的中心点和长宽值
明明只需要log， 却除以了log（2），转换底，因为c++里面的log是默认以e为底，然而e不好打出来，另外的 fit_ratio是为了让值显得大一些，才进行的。
train中
void CHECK_IOU(Anchor
int target_num = GT_Boxes.size();
Dtype area1 = anchor.width * anchor.height;
Dtype maxoverlap = 0.f;
int index = -1;
for (int i = 0; i < target_num; i++) {
const Box
Dtype x0 = std::max(anchor.x_ctr - anchor.width / 2, gt_box.x0);
Dtype y0 = std::max(anchor.y_ctr - anchor.height / 2, gt_box.y0);
Dtype x1 = std::min(anchor.x_ctr + anchor.width / 2, gt_box.x0 + gt_box.width);
Dtype y1 = std::min(anchor.y_ctr + anchor.height / 2, gt_box.y0 + gt_box.height);
Dtype area = (x1 - x0 + 1) * (y1 - y0 + 1);
if (x1 - x0 + 1 < 0 || y1 - y0 + 1 < 0){
continue;
}
Dtype area2 = gt_box.width * gt_box.height;
Dtype Overlap = area / (area1 + area2 - area);
if (Overlap > maxoverlap) {
maxoverlap = Overlap;
index = i;
}
}
anchor.overlap = maxoverlap;
if (maxoverlap > Part_overlap) {
const Box
anchor.name = box.name;
if (maxoverlap > Pos_overlap) {
anchor.isobject = 2;
}
else {
anchor.isobject = 1;
}
anchor.Tar[0] = (box.x0 + box.width / 2 - anchor.x_ctr) / anchor.width;
anchor.Tar[1] = (box.y0 + box.height / 2 - anchor.y_ctr) / anchor.height;
anchor.Tar[2] = std::log(box.width / anchor.width) / std::log(2) * fit_ratio;
anchor.Tar[3] = std::log(box.height / anchor.height) / std::log(2) * fit_ratio;
}
else {
anchor.isobject = 0;
caffe::caffe_memset(4 * sizeof(Dtype), 0, anchor.Tar);
}
if (LabelIndex.count(anchor.name)) {
anchor.cls = LabelIndex.find(anchor.name)->second;
}
}
test中
微调box
template < typename Dtype >
inline void ReassignBox(Box &box, const Anchor
Dtype x_ctr_off, Dtype y_ctr_off, Dtype w_off, Dtype h_off){
box.width = std::pow(2, w_off / fit_ratio) * anchor.width;
box.height = std::pow(2, h_off / fit_ratio) * anchor.height;
box.x0 = x_ctr_off * anchor.width + anchor.x_ctr - box.width / 2;
box.y0 = y_ctr_off * anchor.height + anchor.y_ctr - box.height / 2;
}