https://www.jianshu.com/p/5195165bbd06
1.step_w、step_h其实就相当于faster中的feat_stride,也就是把这些点从feature map映射回原图,同时也可以看出min_size、max_size这些都是直接在针对原图来讲的
2.以mobileNet-ssd为例子:https://github.com/chuanqi305/MobileNet-SSD/blob/master/train.prototxt
layer {
name: "conv11_mbox_priorbox"
type: "PriorBox"
bottom: "conv11"
bottom: "data"
top: "conv11_mbox_priorbox"
prior_box_param {
min_size: 60.0
aspect_ratio: 2.0
flip: true
clip: false
variance: 0.1
variance: 0.1
variance: 0.2
variance: 0.2
offset: 0.5
}
}layer {
name: "conv13_mbox_priorbox"
type: "PriorBox"
bottom: "conv13"
bottom: "data"
top: "conv13_mbox_priorbox"
prior_box_param {
min_size: 105.0
max_size: 150.0
aspect_ratio: 2.0
aspect_ratio: 3.0
flip: true
clip: false
variance: 0.1
variance: 0.1
variance: 0.2
variance: 0.2
offset: 0.5
}
}只有conv11的anchor个数是3,其他5层都是6,原因是conv11只有min_size,没有max_size,并且aspect_ratio只有1个,其他5层都是两个,也就是说conv11是1+1*2=3,其他5层是1+1+2*2=6
prior_box_layer.cpp里,aspect_ratios_根据这层的param存储相应的aspect ratio.如果flip为true,param里一个aspect ratio就要存储他本身和他的倒数两个值
aspect_ratios_.clear();
aspect_ratios_.push_back(.);
flip_ = prior_box_param.flip();
for (int i = ; i < prior_box_param.aspect_ratio_size(); ++i) {
float ar = prior_box_param.aspect_ratio(i);
bool already_exist = false;
for (int j = ; j < aspect_ratios_.size(); ++j) { //检查是否有重复的
if (fabs(ar - aspect_ratios_[j]) < 1e-) {
already_exist = true;
break;
}
}
if (!already_exist) {
aspect_ratios_.push_back(ar); //如果flip为true,存储aspect ratio和他的倒数,否则只存储aspect ratio本身
if (flip_) {
aspect_ratios_.push_back(./ar);
}
}
}
对于每个点,先计算以min_size为长宽的正方形这个anchor;然后如果有max_size,计算以sqrt(min_size_ * max_size_)为长宽的正方形;然后计算aspect_ratios_中所有的aspect ratios,然后以这个aspect ratios计算box_width = min_size_ * sqrt(ar)和box_height = min_size_ / sqrt(ar),prototxt中的param里,一个ratio要存储他和他的倒数,这样一个ratio就要求两个anchor
for (int h = ; h < layer_height; ++h) {
for (int w = ; w < layer_width; ++w) {
float center_x = (w + offset_) * step_w;
float center_y = (h + offset_) * step_h;
float box_width, box_height;
for (int s = ; s < min_sizes_.size(); ++s) {
int min_size_ = min_sizes_[s];
// first prior: aspect_ratio = 1, size = min_size
box_width = box_height = min_size_;
// xmin
top_data[idx++] = (center_x - box_width / .) / img_width;
// ymin
top_data[idx++] = (center_y - box_height / .) / img_height;
// xmax
top_data[idx++] = (center_x + box_width / .) / img_width;
// ymax
top_data[idx++] = (center_y + box_height / .) / img_height;
if (max_sizes_.size() > ) {
CHECK_EQ(min_sizes_.size(), max_sizes_.size());
int max_size_ = max_sizes_[s];
// second prior: aspect_ratio = 1, size = sqrt(min_size * max_size)
box_width = box_height = sqrt(min_size_ * max_size_);
// xmin
top_data[idx++] = (center_x - box_width / .) / img_width;
// ymin
top_data[idx++] = (center_y - box_height / .) / img_height;
// xmax
top_data[idx++] = (center_x + box_width / .) / img_width;
// ymax
top_data[idx++] = (center_y + box_height / .) / img_height;
}
// rest of priors
for (int r = ; r < aspect_ratios_.size(); ++r) {
float ar = aspect_ratios_[r];
if (fabs(ar - .) < 1e-) {
continue;
}
box_width = min_size_ * sqrt(ar);
box_height = min_size_ / sqrt(ar);
// xmin
top_data[idx++] = (center_x - box_width / .) / img_width;
// ymin
top_data[idx++] = (center_y - box_height / .) / img_height;
// xmax
top_data[idx++] = (center_x + box_width / .) / img_width;
// ymax
top_data[idx++] = (center_y + box_height / .) / img_height;
}
}
}
}3.从reshape可以看出,输出的shape是(1,2,layer_width * layer_height * num_priors_ * 4),layer_width * layer_height * num_priors_ * 4是每个feature map上每个点乘以anchor数,再每个anchor乘以对应的4个坐标,比如整个blob中第一个4个值存储的就是feature map中第一个像素点的min size对应的正方形那个anchor的4个坐标值,第二个就是第一个像素点对应的max size对应的anchor的4个坐标值
void PriorBoxLayer<Dtype>::Reshape(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
const int layer_width = bottom[]->width();
const int layer_height = bottom[]->height();
vector<int> top_shape(, );
// Since all images in a batch has same height and width, we only need to
// generate one set of priors which can be shared across all images.
top_shape[] = ;
// 2 channels. First channel stores the mean of each prior coordinate.
// Second channel stores the variance of each prior coordinate.
top_shape[] = ;
top_shape[] = layer_width * layer_height * num_priors_ * ;
CHECK_GT(top_shape[], );
top[]->Reshape(top_shape);
}
注意到,输出是2channel的,第一个channel就是存储的真实的每个anchor的4个坐标,第二个channel存储的就是variance,variance_在layer_setup里面就初始化了4个值,这4个值就是来自于prototxt的param.这4个值分别对应4个坐标点,对于每个anchor,都会有对应这4个variance值,这些值存储在第二个channel,并且在第二个channel里面每4个值每4个值重复
top_data += top[]->offset(, );
if (variance_.size() == ) {
caffe_set<Dtype>(dim, Dtype(variance_[]), top_data);
} else {
int count = ;
for (int h = ; h < layer_height; ++h) {
for (int w = ; w < layer_width; ++w) {
for (int i = ; i < num_priors_; ++i) {
for (int j = ; j < ; ++j) {
top_data[count] = variance_[j];
++count;
}
}
}
}
}
4.http://www.360doc.com/content/17/0810/10/10408243_678091430.shtml
这两段代码都来自于bbox_util.cpp的DecodeBBox函数.prior_box层输出的prior_variance就是一个系数,这个系数乘以bounding box regression的回归值,在faster中,是直接在anchor的坐标上加bounding box regression,ssd这里可以对回归乘以一个系数.当然DecodeBBox其实也可以使用faster那种方式,可以通过参数控制
else {
// variance is encoded in bbox, we need to scale the offset accordingly.
decode_bbox->set_xmin(
prior_bbox.xmin() + prior_variance[] * bbox.xmin());
decode_bbox->set_ymin(
prior_bbox.ymin() + prior_variance[] * bbox.ymin());
decode_bbox->set_xmax(
prior_bbox.xmax() + prior_variance[] * bbox.xmax());
decode_bbox->set_ymax(
prior_bbox.ymax() + prior_variance[] * bbox.ymax());
}else {
// variance is encoded in bbox, we need to scale the offset accordingly.
decode_bbox->set_xmin(
prior_bbox.xmin() + prior_variance[] * bbox.xmin() * prior_width);
decode_bbox->set_ymin(
prior_bbox.ymin() + prior_variance[] * bbox.ymin() * prior_height);
decode_bbox->set_xmax(
prior_bbox.xmax() + prior_variance[] * bbox.xmax() * prior_width);
decode_bbox->set_ymax(
prior_bbox.ymax() + prior_variance[] * bbox.ymax() * prior_height);
}5.https://zhuanlan.zhihu.com/p/33544892 这个介绍了每层的prior如何确定min_size
对于后面的特征图，先验框尺度按照上面公式线性增加，但是先将尺度比例先扩大100倍，此时增长步长为 ，这样各个特征图的
为
，将这些比例除以100，然后再乘以图片大小，可以得到各个特征图的尺度为
，这种计算方式是参考SSD的Caffe源码。综上，可以得到各个特征图的先验框尺度