出自李子青老师组的作品,FaceBoxes: A CPU Real-time Face Detector with High Accuracy
文章整体创新分为3个部分:
(1)RDCL模块
在这个模块中,卷积的滑动步长是很大,属于比较稀疏的滑动卷积。其中conv1滑动步长为4,使得feature map缩小1/4,conv2使得滑动步长为2,使得feature map缩小1/2,pool1,pool2分别缩小1/2,最终feature map缩小1/32。这样就使得feature map减小的比较快,速度也就会提升。
Conv1,conv2使用的卷积核大小分别为7*7,5*5。
另一个需要说的就是,这里使用了CRELU单元,直接对卷积层的输出,做完batchnorm后,将原始的BN的输出和经过取反后的输出做concat,使得channel数double,然后再做scale操作,这里,caffe中,BN和scale是2个分开的层,BN+scale才算是一个完整的batchnorm。
使用CRELU的优势就在于,减少了一半的卷积计算量。
CRELU的思想在于,刚开始的时候,也就是网络的前几层,网络会学习正样本的信息和负样本的信息,或者说attention的地方和not attention的地方,学习出来的特征是呈现对称的一正一负分布的(pair filter)。从而保证正负相关性。一般网络的前几层会表现出正负先关性,随着层数的增加,这种性质越来越弱。而RELU的作用就在于激活attention的地方,将不care的地方至为0。一般网络的前2层可以尝试使用CRELU加速,同时对精度影响较小。
(2)MSCL模块
该模块通过使Inception module作为一个shortcut来进行不同层的特征融合,保证不同scale的特征都能利用起来。
另外就是采用了SSD的多个scale思想,每一个scale部分都输出该阶段的结果,进行分类和回归操作。这样做能保证不同的scale下检测不同尺度的人脸,可以大大提高召回率。
(3)anchor稠密化策略模块
Incetpiton,conv3_2,conv4_2模块的anchor scale分别为(32,64,128),256,512,anchor采样间隔分别为(32,32,32),64,128
根据上面的公式,可以算出anchor密度分别为(1,2,4),4,4,这里由于前面的anchor密度比较少,会导致对小人脸召回率不够高。作者这里做了anchor的平衡,对前面的inception层的anchor进行了增加。
这里anchor增加的策略就是以感受野为中心,通过像素的滑动,进行45度和135度斜对角线anchor的增加。
代码测试:
git clone https://github.com/weiliu89/caffe.git
git checkout ssd
git pull
git clone https://github.com/zeusees/FaceBoxes.git
下载好模型后使用下面的c++程序进行测试
#define USE_OPENCV 1
//#define CPU_ONLY 1
#define USE_CUDNN 1
#include <caffe/caffe.hpp>
#ifdef USE_OPENCV
#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>
#endif // USE_OPENCV
#include <algorithm>
#include <iomanip>
#include <iosfwd>
#include <memory>
#include <string>
#include <utility>
#include <vector>
#include "head.h"
#ifdef USE_OPENCV
using namespace caffe; // NOLINT(build/namespaces)
class Detector {
public:
Detector(const string& model_file,
const string& weights_file,
const string& mean_file,
const string& mean_value);
std::vector<vector<float> > Detect(const cv::Mat& img);
private:
void SetMean(const string& mean_file, const string& mean_value);
void WrapInputLayer(std::vector<cv::Mat>* input_channels);
void Preprocess(const cv::Mat& img,
std::vector<cv::Mat>* input_channels);
private:
shared_ptr<Net<float> > net_;
cv::Size input_geometry_;
int num_channels_;
cv::Mat mean_;
};
Detector::Detector(const string& model_file,
const string& weights_file,
const string& mean_file,
const string& mean_value) {
#ifdef CPU_ONLY
Caffe::set_mode(Caffe::CPU);
#else
Caffe::set_mode(Caffe::GPU);
#endif
/* Load the network. */
net_.reset(new Net<float>(model_file, TEST));
net_->CopyTrainedLayersFrom(weights_file);
CHECK_EQ(net_->num_inputs(), 1) << "Network should have exactly one input.";
CHECK_EQ(net_->num_outputs(), 1) << "Network should have exactly one output.";
Blob<float>* input_layer = net_->input_blobs()[0];
num_channels_ = input_layer->channels();
CHECK(num_channels_ == 3 || num_channels_ == 1)
<< "Input layer should have 1 or 3 channels.";
input_geometry_ = cv::Size(input_layer->width(), input_layer->height());
/* Load the binaryproto mean file. */
SetMean(mean_file, mean_value);
}
std::vector<vector<float> > Detector::Detect(const cv::Mat& img) {
Blob<float>* input_layer = net_->input_blobs()[0];
input_layer->Reshape(1, num_channels_,
input_geometry_.height, input_geometry_.width);
/* Forward dimension change to all layers. */
net_->Reshape();
std::vector<cv::Mat> input_channels;
WrapInputLayer(&input_channels);
Preprocess(img, &input_channels);
net_->Forward();
/* Copy the output layer to a std::vector */
Blob<float>* result_blob = net_->output_blobs()[0];
const float* result = result_blob->cpu_data();
const int num_det = result_blob->height();
vector<vector<float> > detections;
for (int k = 0; k < num_det; ++k) {
if (result[0] == -1) {
// Skip invalid detection.
result += 7;
continue;
}
vector<float> detection(result, result + 7);
detections.push_back(detection);
result += 7;
}
return detections;
}
/* Load the mean file in binaryproto format. */
void Detector::SetMean(const string& mean_file, const string& mean_value) {
cv::Scalar channel_mean;
if (!mean_file.empty()) {
CHECK(mean_value.empty()) <<
"Cannot specify mean_file and mean_value at the same time";
BlobProto blob_proto;
ReadProtoFromBinaryFileOrDie(mean_file.c_str(), &blob_proto);
/* Convert from BlobProto to Blob<float> */
Blob<float> mean_blob;
mean_blob.FromProto(blob_proto);
CHECK_EQ(mean_blob.channels(), num_channels_)
<< "Number of channels of mean file doesn't match input layer.";
/* The format of the mean file is planar 32-bit float BGR or grayscale. */
std::vector<cv::Mat> channels;
float* data = mean_blob.mutable_cpu_data();
for (int i = 0; i < num_channels_; ++i) {
/* Extract an individual channel. */
cv::Mat channel(mean_blob.height(), mean_blob.width(), CV_32FC1, data);
channels.push_back(channel);
data += mean_blob.height() * mean_blob.width();
}
/* Merge the separate channels into a single image. */
cv::Mat mean;
cv::merge(channels, mean);
/* Compute the global mean pixel value and create a mean image
* filled with this value. */
channel_mean = cv::mean(mean);
mean_ = cv::Mat(input_geometry_, mean.type(), channel_mean);
}
if (!mean_value.empty()) {
CHECK(mean_file.empty()) <<
"Cannot specify mean_file and mean_value at the same time";
stringstream ss(mean_value);
vector<float> values;
string item;
while (getline(ss, item, ',')) {
float value = std::atof(item.c_str());
values.push_back(value);
}
CHECK(values.size() == 1 || values.size() == num_channels_) <<
"Specify either 1 mean_value or as many as channels: " << num_channels_;
std::vector<cv::Mat> channels;
for (int i = 0; i < num_channels_; ++i) {
/* Extract an individual channel. */
cv::Mat channel(input_geometry_.height, input_geometry_.width, CV_32FC1,
cv::Scalar(values[i]));
channels.push_back(channel);
}
cv::merge(channels, mean_);
}
}
/* Wrap the input layer of the network in separate cv::Mat objects
* (one per channel). This way we save one memcpy operation and we
* don't need to rely on cudaMemcpy2D. The last preprocessing
* operation will write the separate channels directly to the input
* layer. */
void Detector::WrapInputLayer(std::vector<cv::Mat>* input_channels) {
Blob<float>* input_layer = net_->input_blobs()[0];
int width = input_layer->width();
int height = input_layer->height();
float* input_data = input_layer->mutable_cpu_data();
for (int i = 0; i < input_layer->channels(); ++i) {
cv::Mat channel(height, width, CV_32FC1, input_data);
input_channels->push_back(channel);
input_data += width * height;
}
}
void Detector::Preprocess(const cv::Mat& img,
std::vector<cv::Mat>* input_channels) {
/* Convert the input image to the input image format of the network. */
cv::Mat sample;
if (img.channels() == 3 && num_channels_ == 1)
cv::cvtColor(img, sample, cv::COLOR_BGR2GRAY);
else if (img.channels() == 4 && num_channels_ == 1)
cv::cvtColor(img, sample, cv::COLOR_BGRA2GRAY);
else if (img.channels() == 4 && num_channels_ == 3)
cv::cvtColor(img, sample, cv::COLOR_BGRA2BGR);
else if (img.channels() == 1 && num_channels_ == 3)
cv::cvtColor(img, sample, cv::COLOR_GRAY2BGR);
else
sample = img;
cv::Mat sample_resized;
if (sample.size() != input_geometry_)
cv::resize(sample, sample_resized, input_geometry_);
else
sample_resized = sample;
cv::Mat sample_float;
if (num_channels_ == 3)
sample_resized.convertTo(sample_float, CV_32FC3);
else
sample_resized.convertTo(sample_float, CV_32FC1);
cv::Mat sample_normalized;
cv::subtract(sample_float, mean_, sample_normalized);
sample_normalized = sample_normalized / 127.5;
/* This operation will write the separate BGR planes directly to the
* input layer of the network because it is wrapped by the cv::Mat
* objects in input_channels. */
cv::split(sample_normalized, *input_channels);
CHECK(reinterpret_cast<float*>(input_channels->at(0).data)
== net_->input_blobs()[0]->cpu_data())
<< "Input channels are not wrapping the input layer of the network.";
}
DEFINE_string(mean_file, "",
"The mean file used to subtract from the input image.");
DEFINE_string(mean_value, "104,117,123",
"If specified, can be one value or can be same as image channels"
" - would subtract from the corresponding channel). Separated by ','."
"Either mean_file or mean_value should be provided, not both.");
DEFINE_string(file_type, "image",
"The file type in the list_file. Currently support image and video.");
DEFINE_string(out_file, "",
"If provided, store the detection results in the out_file.");
DEFINE_double(confidence_threshold, 0.3,//0.01,
"Only store detections with score higher than the threshold.");
int main(int argc, char** argv) {
::google::InitGoogleLogging(argv[0]);
argc = 4;
argv[0] = "ssd_detect.exe -file_type=image -confidence_threshold=0.3";
argv[1] = "faceboxes_deploy.prototxt";
argv[2] = "FaceBoxes_1024x1024.caffemodel";
argv[3] = "list.txt";
// ::google::InitGoogleLogging(argv[0]);
// Print output to stderr (while still logging)
FLAGS_alsologtostderr = 1;
#ifndef GFLAGS_GFLAGS_H_
namespace gflags = google;
#endif
gflags::SetUsageMessage("Do detection using SSD mode.\n"
"Usage:\n"
" ssd_detect [FLAGS] model_file weights_file list_file\n");
gflags::ParseCommandLineFlags(&argc, &argv, true);
if (argc < 4) {
gflags::ShowUsageWithFlagsRestrict(argv[0], "examples/ssd/ssd_detect");
return 1;
}
const string& model_file = argv[1];
const string& weights_file = argv[2];
const string& mean_file = FLAGS_mean_file;
const string& mean_value = FLAGS_mean_value;
const string& file_type = FLAGS_file_type;
const string& out_file = FLAGS_out_file;
const float confidence_threshold = FLAGS_confidence_threshold;
// Initialize the network.
Detector detector(model_file, weights_file, mean_file, mean_value);
// Set the output mode.
std::streambuf* buf = std::cout.rdbuf();
std::ofstream outfile;
if (!out_file.empty()) {
outfile.open(out_file.c_str());
if (outfile.good()) {
buf = outfile.rdbuf();
}
}
std::ostream out(buf);
// Process image one by one.
std::ifstream infile(argv[3]);
std::string file;
while (infile >> file) {
if (file_type == "image") {
cv::Mat img = cv::imread(file, -1);
CHECK(!img.empty()) << "Unable to decode image " << file;
clock_t start1, end1;
start1 = clock();
std::vector<vector<float> > detections = detector.Detect(img);
end1 = (double)(1000 * (clock() - start1) / CLOCKS_PER_SEC);
std::cout << "time: " << end1 << "ms" << std::endl;
/* Print the detection results. */
for (int i = 0; i < detections.size(); ++i) {
const vector<float>& d = detections[i];
// Detection format: [image_id, label, score, xmin, ymin, xmax, ymax].
CHECK_EQ(d.size(), 7);
const float score = d[2];
if (score >= confidence_threshold) {
out << file << " ";
out << static_cast<int>(d[1]) << " ";
out << score << " ";
out << static_cast<int>(d[3] * img.cols) << " ";
out << static_cast<int>(d[4] * img.rows) << " ";
out << static_cast<int>(d[5] * img.cols) << " ";
out << static_cast<int>(d[6] * img.rows) << std::endl;
cv::rectangle( img, cv::Point(static_cast<int>(d[3] * img.cols), static_cast<int>(d[4] * img.rows)), cv::Point(static_cast<int>(d[5] * img.cols), static_cast<int>(d[6] * img.rows)), cv::Scalar(0, static_cast<int>(d[1])*10, 205), 2, 4, 0 );
}
}
imshow("result",img);
cvWaitKey(0);
} else if (file_type == "video") {
cv::VideoCapture cap(file);
if (!cap.isOpened()) {
LOG(FATAL) << "Failed to open video: " << file;
}
cv::Mat img;
int frame_count = 0;
while (true) {
bool success = cap.read(img);
if (!success) {
LOG(INFO) << "Process " << frame_count << " frames from " << file;
break;
}
CHECK(!img.empty()) << "Error when read frame";
clock_t start1, end1;
start1 = clock();
std::vector<vector<float> > detections = detector.Detect(img);
end1 = (double)(1000 * (clock() - start1) / CLOCKS_PER_SEC);
std::cout << "time: " << end1 << "ms" << std::endl;
/* Print the detection results. */
for (int i = 0; i < detections.size(); ++i) {
const vector<float>& d = detections[i];
// Detection format: [image_id, label, score, xmin, ymin, xmax, ymax].
CHECK_EQ(d.size(), 7);
const float score = d[2];
if (score >= confidence_threshold) {
out << file << "_";
out << std::setfill('0') << std::setw(6) << frame_count << " ";
out << static_cast<int>(d[1]) << " ";
out << score << " ";
out << static_cast<int>(d[3] * img.cols) << " ";
out << static_cast<int>(d[4] * img.rows) << " ";
out << static_cast<int>(d[5] * img.cols) << " ";
out << static_cast<int>(d[6] * img.rows) << std::endl;
cv::rectangle( img, cv::Point(static_cast<int>(d[3] * img.cols), static_cast<int>(d[4] * img.rows)), cv::Point(static_cast<int>(d[5] * img.cols), static_cast<int>(d[6] * img.rows)), cv::Scalar(0, static_cast<int>(d[1])*10, 255), 2, 4, 0 );
}
}
++frame_count;
imshow("result",img);
cvWaitKey(1);
}
if (cap.isOpened()) {
cap.release();
}
} else {
LOG(FATAL) << "Unknown file_type: " << file_type;
}
}
return 0;
}
#else
int main(int argc, char** argv) {
LOG(FATAL) << "This example requires OpenCV; compile with USE_OPENCV.";
}
#endif // USE_OPENCV
输出结果:(cpu 7700HQ,显卡1060)
结果看速度确实很快,模型也很小,只有3.4M,召回率不是很好,可能作者的模型实现上和论文的实现有区别吧。留待后续探索。
my test project:http://download.csdn.net/download/qq_14845119/10216589
References:
FaceBoxes: A CPU Real-time Face Detector with HighAccuracy
Understanding and improving convolutional neural networks viaconcatenated rectified linear units. InICML, 2016
https://github.com/zeusees/FaceBoxes
官方github: