From 1b5afb45838e603fa6780762eb8cc59246dc2d81 Mon Sep 17 00:00:00 2001
From: IlyaOvodov <b@ovdv.ru>
Date: Tue, 08 May 2018 11:09:35 +0000
Subject: [PATCH] Output improvements for detector results: When printing detector results, output was done in random order, obfuscating results for interpreting. Now: 1. Text output includes coordinates of rects in (left,right,top,bottom in pixels) along with label and score 2. Text output is sorted by rect lefts to simplify finding appropriate rects on image 3. If several class probs are > thresh for some detection, the most probable is written first and coordinates for others are not repeated 4. Rects are imprinted in image in order by their best class prob, so most probable rects are always on top and not overlayed by less probable ones 5. Most probable label for rect is always written first Also: 6. Message about low GPU memory include required amount
---
src/yolo_console_dll.cpp | 324 +++++++++++++++++++++++++++++++++--------------------
1 files changed, 202 insertions(+), 122 deletions(-)
diff --git a/src/yolo_console_dll.cpp b/src/yolo_console_dll.cpp
index 1e9ba98..ea4330e 100644
--- a/src/yolo_console_dll.cpp
+++ b/src/yolo_console_dll.cpp
@@ -11,11 +11,16 @@
#ifdef _WIN32
#define OPENCV
+#define GPU
#endif
// To use tracking - uncomment the following line. Tracking is supported only by OpenCV 3.x
//#define TRACK_OPTFLOW
+//#include "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.1\include\cuda_runtime.h"
+//#pragma comment(lib, "C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v9.1/lib/x64/cudart.lib")
+//static std::shared_ptr<image_t> device_ptr(NULL, [](void *img) { cudaDeviceReset(); });
+
#include "yolo_v2_class.hpp" // imported functions from DLL
#ifdef OPENCV
@@ -25,135 +30,168 @@
#include "opencv2/videoio/videoio.hpp"
#define OPENCV_VERSION CVAUX_STR(CV_VERSION_MAJOR)""CVAUX_STR(CV_VERSION_MINOR)""CVAUX_STR(CV_VERSION_REVISION)
#pragma comment(lib, "opencv_world" OPENCV_VERSION ".lib")
+#ifdef TRACK_OPTFLOW
#pragma comment(lib, "opencv_cudaoptflow" OPENCV_VERSION ".lib")
#pragma comment(lib, "opencv_cudaimgproc" OPENCV_VERSION ".lib")
#pragma comment(lib, "opencv_core" OPENCV_VERSION ".lib")
#pragma comment(lib, "opencv_imgproc" OPENCV_VERSION ".lib")
#pragma comment(lib, "opencv_highgui" OPENCV_VERSION ".lib")
+#endif // TRACK_OPTFLOW
#else
#define OPENCV_VERSION CVAUX_STR(CV_VERSION_EPOCH)""CVAUX_STR(CV_VERSION_MAJOR)""CVAUX_STR(CV_VERSION_MINOR)
#pragma comment(lib, "opencv_core" OPENCV_VERSION ".lib")
#pragma comment(lib, "opencv_imgproc" OPENCV_VERSION ".lib")
#pragma comment(lib, "opencv_highgui" OPENCV_VERSION ".lib")
-#endif
+#endif // CV_VERSION_EPOCH
-
-cv::Scalar obj_id_to_color(int obj_id) {
- int const colors[6][3] = { { 1,0,1 },{ 0,0,1 },{ 0,1,1 },{ 0,1,0 },{ 1,1,0 },{ 1,0,0 } };
- int const offset = obj_id * 123457 % 6;
- int const color_scale = 150 + (obj_id * 123457) % 100;
- cv::Scalar color(colors[offset][0], colors[offset][1], colors[offset][2]);
- color *= color_scale;
- return color;
-}
-
-class preview_boxes_t {
- enum { frames_history = 30 }; // how long to keep the history saved
-
- struct preview_box_track_t {
- unsigned int track_id, obj_id, last_showed_frames_ago;
- cv::Mat mat_obj;
- preview_box_track_t() : track_id(0), obj_id(0), last_showed_frames_ago(frames_history) {}
- };
- std::vector<preview_box_track_t> preview_box_track_id;
- size_t const preview_box_size, const bottom_offset;
- bool const one_off_detections;
+class track_kalman {
public:
- preview_boxes_t(size_t _preview_box_size = 100, size_t _bottom_offset = 100, bool _one_off_detections = false) :
- preview_box_size(_preview_box_size), bottom_offset(_bottom_offset), one_off_detections(_one_off_detections)
- {}
+ cv::KalmanFilter kf;
+ int state_size, meas_size, contr_size;
- void draw_preview_boxes(cv::Mat src_mat, cv::Mat draw_mat, std::vector<bbox_t> result_vec, bool draw_boxes = true)
+
+ track_kalman(int _state_size = 10, int _meas_size = 10, int _contr_size = 0)
+ : state_size(_state_size), meas_size(_meas_size), contr_size(_contr_size)
{
- size_t const count_preview_boxes = draw_mat.cols / preview_box_size;
- if (preview_box_track_id.size() != count_preview_boxes) preview_box_track_id.resize(count_preview_boxes);
+ kf.init(state_size, meas_size, contr_size, CV_32F);
- // increment frames history
- for (auto &i : preview_box_track_id)
- i.last_showed_frames_ago = std::min((unsigned)frames_history, i.last_showed_frames_ago + 1);
+ cv::setIdentity(kf.measurementMatrix);
+ cv::setIdentity(kf.measurementNoiseCov, cv::Scalar::all(1e-1));
+ cv::setIdentity(kf.processNoiseCov, cv::Scalar::all(1e-5));
+ cv::setIdentity(kf.errorCovPost, cv::Scalar::all(1e-2));
+ cv::setIdentity(kf.transitionMatrix);
+ }
- // occupy empty boxes
- for (auto &k : result_vec) {
- bool found = false;
- for (auto &i : preview_box_track_id) {
- if (i.track_id == k.track_id) {
- if(!one_off_detections)
- i.last_showed_frames_ago = 0;
- found = true;
- break;
- }
- }
- if (!found) {
- for (auto &i : preview_box_track_id) {
- if (i.last_showed_frames_ago == frames_history) {
- if (!one_off_detections && k.frames_counter == 0) break;
- i.track_id = k.track_id;
- i.obj_id = k.obj_id;
- i.last_showed_frames_ago = 0;
- break;
- }
- }
- }
+ void set(std::vector<bbox_t> result_vec) {
+ for (size_t i = 0; i < result_vec.size() && i < state_size*2; ++i) {
+ kf.statePost.at<float>(i * 2 + 0) = result_vec[i].x;
+ kf.statePost.at<float>(i * 2 + 1) = result_vec[i].y;
}
+ }
- // draw preview box (from old or current frame)
- for (size_t i = 0; i < preview_box_track_id.size(); ++i)
- {
- // get object image
- cv::Mat dst = preview_box_track_id[i].mat_obj;
- bool current_detection = false;
+ // Kalman.correct() calculates: statePost = statePre + gain * (z(k)-measurementMatrix*statePre);
+ // corrected state (x(k)): x(k)=x'(k)+K(k)*(z(k)-H*x'(k))
+ std::vector<bbox_t> correct(std::vector<bbox_t> result_vec) {
+ cv::Mat measurement(meas_size, 1, CV_32F);
+ for (size_t i = 0; i < result_vec.size() && i < meas_size * 2; ++i) {
+ measurement.at<float>(i * 2 + 0) = result_vec[i].x;
+ measurement.at<float>(i * 2 + 1) = result_vec[i].y;
+ }
+ cv::Mat estimated = kf.correct(measurement);
+ for (size_t i = 0; i < result_vec.size() && i < meas_size * 2; ++i) {
+ result_vec[i].x = estimated.at<float>(i * 2 + 0);
+ result_vec[i].y = estimated.at<float>(i * 2 + 1);
+ }
+ return result_vec;
+ }
- for (auto &k : result_vec) {
- if (preview_box_track_id[i].track_id == k.track_id) {
- if (one_off_detections && preview_box_track_id[i].last_showed_frames_ago > 0) break;
- bbox_t b = k;
- cv::Rect r(b.x, b.y, b.w, b.h);
- cv::Rect img_rect(cv::Point2i(0, 0), src_mat.size());
- cv::Rect rect_roi = r & img_rect;
- if (rect_roi.width > 1 || rect_roi.height > 1) {
- cv::Mat roi = src_mat(rect_roi);
- cv::resize(roi, dst, cv::Size(preview_box_size, preview_box_size));
- preview_box_track_id[i].mat_obj = dst.clone();
- current_detection = true;
+ // Kalman.predict() calculates: statePre = TransitionMatrix * statePost;
+ // predicted state (x'(k)): x(k)=A*x(k-1)+B*u(k)
+ std::vector<bbox_t> predict() {
+ std::vector<bbox_t> result_vec;
+ cv::Mat control;
+ cv::Mat prediction = kf.predict(control);
+ for (size_t i = 0; i < prediction.rows && i < state_size * 2; ++i) {
+ result_vec[i].x = prediction.at<float>(i * 2 + 0);
+ result_vec[i].y = prediction.at<float>(i * 2 + 1);
+ }
+ return result_vec;
+ }
+
+};
+
+
+
+
+class extrapolate_coords_t {
+public:
+ std::vector<bbox_t> old_result_vec;
+ std::vector<float> dx_vec, dy_vec, time_vec;
+ std::vector<float> old_dx_vec, old_dy_vec;
+
+ void new_result(std::vector<bbox_t> new_result_vec, float new_time) {
+ old_dx_vec = dx_vec;
+ old_dy_vec = dy_vec;
+ if (old_dx_vec.size() != old_result_vec.size()) std::cout << "old_dx != old_res \n";
+ dx_vec = std::vector<float>(new_result_vec.size(), 0);
+ dy_vec = std::vector<float>(new_result_vec.size(), 0);
+ update_result(new_result_vec, new_time, false);
+ old_result_vec = new_result_vec;
+ time_vec = std::vector<float>(new_result_vec.size(), new_time);
+ }
+
+ void update_result(std::vector<bbox_t> new_result_vec, float new_time, bool update = true) {
+ for (size_t i = 0; i < new_result_vec.size(); ++i) {
+ for (size_t k = 0; k < old_result_vec.size(); ++k) {
+ if (old_result_vec[k].track_id == new_result_vec[i].track_id && old_result_vec[k].obj_id == new_result_vec[i].obj_id) {
+ float const delta_time = new_time - time_vec[k];
+ if (abs(delta_time) < 1) break;
+ size_t index = (update) ? k : i;
+ float dx = ((float)new_result_vec[i].x - (float)old_result_vec[k].x) / delta_time;
+ float dy = ((float)new_result_vec[i].y - (float)old_result_vec[k].y) / delta_time;
+ float old_dx = dx, old_dy = dy;
+
+ // if it's shaking
+ if (update) {
+ if (dx * dx_vec[i] < 0) dx = dx / 2;
+ if (dy * dy_vec[i] < 0) dy = dy / 2;
+ } else {
+ if (dx * old_dx_vec[k] < 0) dx = dx / 2;
+ if (dy * old_dy_vec[k] < 0) dy = dy / 2;
}
+ dx_vec[index] = dx;
+ dy_vec[index] = dy;
+
+ //if (old_dx == dx && old_dy == dy) std::cout << "not shakin \n";
+ //else std::cout << "shakin \n";
+
+ if (dx_vec[index] > 1000 || dy_vec[index] > 1000) {
+ //std::cout << "!!! bad dx or dy, dx = " << dx_vec[index] << ", dy = " << dy_vec[index] <<
+ // ", delta_time = " << delta_time << ", update = " << update << std::endl;
+ dx_vec[index] = 0;
+ dy_vec[index] = 0;
+ }
+ old_result_vec[k].x = new_result_vec[i].x;
+ old_result_vec[k].y = new_result_vec[i].y;
+ time_vec[k] = new_time;
break;
}
}
-
- // draw object image
- if (draw_boxes) {
- cv::Mat dst = preview_box_track_id[i].mat_obj;
- if (preview_box_track_id[i].last_showed_frames_ago < frames_history &&
- dst.size() == cv::Size(preview_box_size, preview_box_size))
- {
- cv::Rect dst_rect_roi(cv::Point2i(i * preview_box_size, draw_mat.rows - bottom_offset), dst.size());
- cv::Mat dst_roi = draw_mat(dst_rect_roi);
- dst.copyTo(dst_roi);
-
- cv::Scalar color = obj_id_to_color(preview_box_track_id[i].obj_id);
- int thickness = (current_detection) ? 5 : 1;
- cv::rectangle(draw_mat, dst_rect_roi, color, thickness);
- }
- }
}
}
+
+ std::vector<bbox_t> predict(float cur_time) {
+ std::vector<bbox_t> result_vec = old_result_vec;
+ for (size_t i = 0; i < old_result_vec.size(); ++i) {
+ float const delta_time = cur_time - time_vec[i];
+ auto &bbox = result_vec[i];
+ float new_x = (float) bbox.x + dx_vec[i] * delta_time;
+ float new_y = (float) bbox.y + dy_vec[i] * delta_time;
+ if (new_x > 0) bbox.x = new_x;
+ else bbox.x = 0;
+ if (new_y > 0) bbox.y = new_y;
+ else bbox.y = 0;
+ }
+ return result_vec;
+ }
+
};
void draw_boxes(cv::Mat mat_img, std::vector<bbox_t> result_vec, std::vector<std::string> obj_names,
- unsigned int wait_msec = 0, int current_det_fps = -1, int current_cap_fps = -1)
+ int current_det_fps = -1, int current_cap_fps = -1)
{
int const colors[6][3] = { { 1,0,1 },{ 0,0,1 },{ 0,1,1 },{ 0,1,0 },{ 1,1,0 },{ 1,0,0 } };
for (auto &i : result_vec) {
cv::Scalar color = obj_id_to_color(i.obj_id);
- cv::rectangle(mat_img, cv::Rect(i.x, i.y, i.w, i.h), color, 5);
+ cv::rectangle(mat_img, cv::Rect(i.x, i.y, i.w, i.h), color, 2);
if (obj_names.size() > i.obj_id) {
std::string obj_name = obj_names[i.obj_id];
if (i.track_id > 0) obj_name += " - " + std::to_string(i.track_id);
cv::Size const text_size = getTextSize(obj_name, cv::FONT_HERSHEY_COMPLEX_SMALL, 1.2, 2, 0);
int const max_width = (text_size.width > i.w + 2) ? text_size.width : (i.w + 2);
- cv::rectangle(mat_img, cv::Point2f(std::max((int)i.x - 3, 0), std::max((int)i.y - 30, 0)),
+ cv::rectangle(mat_img, cv::Point2f(std::max((int)i.x - 1, 0), std::max((int)i.y - 30, 0)),
cv::Point2f(std::min((int)i.x + max_width, mat_img.cols-1), std::min((int)i.y, mat_img.rows-1)),
color, CV_FILLED, 8, 0);
putText(mat_img, obj_name, cv::Point2f(i.x, i.y - 10), cv::FONT_HERSHEY_COMPLEX_SMALL, 1.2, cv::Scalar(0, 0, 0), 2);
@@ -163,9 +201,6 @@
std::string fps_str = "FPS detection: " + std::to_string(current_det_fps) + " FPS capture: " + std::to_string(current_cap_fps);
putText(mat_img, fps_str, cv::Point2f(10, 20), cv::FONT_HERSHEY_COMPLEX_SMALL, 1.2, cv::Scalar(50, 255, 0), 2);
}
-
- cv::imshow("window name", mat_img);
- cv::waitKey(wait_msec);
}
#endif // OPENCV
@@ -191,9 +226,9 @@
int main(int argc, char *argv[])
{
- std::string names_file = "data/voc.names";
- std::string cfg_file = "cfg/yolo-voc.cfg";
- std::string weights_file = "yolo-voc.weights";
+ std::string names_file = "data/coco.names";
+ std::string cfg_file = "cfg/yolov3.cfg";
+ std::string weights_file = "yolov3.weights";
std::string filename;
if (argc > 4) { //voc.names yolo-voc.cfg yolo-voc.weights test.mp4
@@ -204,11 +239,13 @@
}
else if (argc > 1) filename = argv[1];
+ float const thresh = (argc > 5) ? std::stof(argv[5]) : 0.20;
+
Detector detector(cfg_file, weights_file);
auto obj_names = objects_names_from_file(names_file);
std::string out_videofile = "result.avi";
- bool const save_output_videofile = false;
+ bool const save_output_videofile = true;
#ifdef TRACK_OPTFLOW
Tracker_optflow tracker_flow;
detector.wait_stream = true;
@@ -222,7 +259,11 @@
try {
#ifdef OPENCV
+ extrapolate_coords_t extrapolate_coords;
+ bool extrapolate_flag = false;
+ float cur_time_extrapolate = 0, old_time_extrapolate = 0;
preview_boxes_t large_preview(100, 150, false), small_preview(50, 50, true);
+ bool show_small_boxes = false;
std::string const file_ext = filename.substr(filename.find_last_of(".") + 1);
std::string const protocol = filename.substr(0, 7);
@@ -236,6 +277,7 @@
std::vector<bbox_t> result_vec, thread_result_vec;
detector.nms = 0.02; // comment it - if track_id is not required
std::atomic<bool> consumed, videowrite_ready;
+ bool exit_flag = false;
consumed = true;
videowrite_ready = true;
std::atomic<int> fps_det_counter, fps_cap_counter;
@@ -262,33 +304,41 @@
cur_frame = cap_frame.clone();
}
t_cap = std::thread([&]() { cap >> cap_frame; });
+ ++cur_time_extrapolate;
// swap result bouned-boxes and input-frame
if(consumed)
{
std::unique_lock<std::mutex> lock(mtx);
det_image = detector.mat_to_image_resize(cur_frame);
- auto old_result_vec = result_vec;
- result_vec = thread_result_vec;
- result_vec = detector.tracking_id(result_vec); // comment it - if track_id is not required
+ auto old_result_vec = detector.tracking_id(result_vec);
+ auto detected_result_vec = thread_result_vec;
+ result_vec = detected_result_vec;
#ifdef TRACK_OPTFLOW
// track optical flow
if (track_optflow_queue.size() > 0) {
- small_preview.draw_preview_boxes(track_optflow_queue.front(), cur_frame, result_vec, false);
-
- std::queue<cv::Mat> new_track_optflow_queue;
//std::cout << "\n !!!! all = " << track_optflow_queue.size() << ", cur = " << passed_flow_frames << std::endl;
- tracker_flow.update_tracking_flow(track_optflow_queue.front());
+ cv::Mat first_frame = track_optflow_queue.front();
+ tracker_flow.update_tracking_flow(track_optflow_queue.front(), result_vec);
while (track_optflow_queue.size() > 1) {
track_optflow_queue.pop();
- result_vec = tracker_flow.tracking_flow(track_optflow_queue.front(), result_vec);
- if (track_optflow_queue.size() <= passed_flow_frames && new_track_optflow_queue.size() == 0)
- new_track_optflow_queue = track_optflow_queue;
+ result_vec = tracker_flow.tracking_flow(track_optflow_queue.front(), true);
}
- track_optflow_queue = new_track_optflow_queue;
+ track_optflow_queue.pop();
passed_flow_frames = 0;
+
+ result_vec = detector.tracking_id(result_vec);
+ auto tmp_result_vec = detector.tracking_id(detected_result_vec, false);
+ small_preview.set(first_frame, tmp_result_vec);
+
+ extrapolate_coords.new_result(tmp_result_vec, old_time_extrapolate);
+ old_time_extrapolate = cur_time_extrapolate;
+ extrapolate_coords.update_result(result_vec, cur_time_extrapolate - 1);
}
+#else
+ result_vec = detector.tracking_id(result_vec); // comment it - if track_id is not required
+ extrapolate_coords.new_result(result_vec, cur_time_extrapolate - 1);
#endif
// add old tracked objects
for (auto &i : old_result_vec) {
@@ -299,10 +349,14 @@
if (i.frames_counter-- > 1)
result_vec.push_back(i);
}
- else
+ else {
it->frames_counter = std::min((unsigned)3, i.frames_counter + 1);
+ }
}
-
+#ifdef TRACK_OPTFLOW
+ tracker_flow.update_cur_bbox_vec(result_vec);
+ result_vec = tracker_flow.tracking_flow(cur_frame, true); // track optical flow
+#endif
consumed = false;
cv_pre_tracked.notify_all();
}
@@ -311,19 +365,22 @@
t_detect = std::thread([&]() {
auto current_image = det_image;
consumed = true;
- while (current_image.use_count() > 0) {
- auto result = detector.detect_resized(*current_image, frame_size, 0.20, false); // true
+ while (current_image.use_count() > 0 && !exit_flag) {
+ auto result = detector.detect_resized(*current_image, frame_size.width, frame_size.height,
+ thresh, false); // true
++fps_det_counter;
std::unique_lock<std::mutex> lock(mtx);
thread_result_vec = result;
- current_image = det_image;
consumed = true;
cv_detected.notify_all();
- if(detector.wait_stream)
- while (consumed) cv_pre_tracked.wait(lock);
+ if (detector.wait_stream) {
+ while (consumed && !exit_flag) cv_pre_tracked.wait(lock);
+ }
+ current_image = det_image;
}
});
}
+ //while (!consumed); // sync detection
if (!cur_frame.empty()) {
steady_end = std::chrono::steady_clock::now();
@@ -335,16 +392,29 @@
fps_cap_counter = 0;
}
+ large_preview.set(cur_frame, result_vec);
#ifdef TRACK_OPTFLOW
++passed_flow_frames;
track_optflow_queue.push(cur_frame.clone());
- result_vec = tracker_flow.tracking_flow(cur_frame, result_vec, true); // track optical flow
- small_preview.draw_preview_boxes(cur_frame.clone(), cur_frame, result_vec, true);
-#endif
- large_preview.draw_preview_boxes(cur_frame.clone(), cur_frame, result_vec, true);
-
- draw_boxes(cur_frame, result_vec, obj_names, 3, current_det_fps, current_cap_fps); // 3 or 16ms
+ result_vec = tracker_flow.tracking_flow(cur_frame); // track optical flow
+ extrapolate_coords.update_result(result_vec, cur_time_extrapolate);
+ small_preview.draw(cur_frame, show_small_boxes);
+#endif
+ auto result_vec_draw = result_vec;
+ if (extrapolate_flag) {
+ result_vec_draw = extrapolate_coords.predict(cur_time_extrapolate);
+ cv::putText(cur_frame, "extrapolate", cv::Point2f(10, 40), cv::FONT_HERSHEY_COMPLEX_SMALL, 1.0, cv::Scalar(50, 50, 0), 2);
+ }
+ draw_boxes(cur_frame, result_vec_draw, obj_names, current_det_fps, current_cap_fps);
//show_console_result(result_vec, obj_names);
+ large_preview.draw(cur_frame);
+
+ cv::imshow("window name", cur_frame);
+ int key = cv::waitKey(3); // 3 or 16ms
+ if (key == 'f') show_small_boxes = !show_small_boxes;
+ if (key == 'p') while (true) if(cv::waitKey(100) == 'p') break;
+ if (key == 'e') extrapolate_flag = !extrapolate_flag;
+ if (key == 27) { exit_flag = true; break; }
if (output_video.isOpened() && videowrite_ready) {
if (t_videowrite.joinable()) t_videowrite.join();
@@ -364,10 +434,12 @@
}
#endif
}
+ exit_flag = true;
if (t_cap.joinable()) t_cap.join();
if (t_detect.joinable()) t_detect.join();
if (t_videowrite.joinable()) t_videowrite.join();
std::cout << "Video ended \n";
+ break;
}
else if (file_ext == "txt") { // list of image files
std::ifstream file(filename);
@@ -385,10 +457,18 @@
}
else { // image file
cv::Mat mat_img = cv::imread(filename);
+
+ auto start = std::chrono::steady_clock::now();
std::vector<bbox_t> result_vec = detector.detect(mat_img);
- result_vec = detector.tracking_id(result_vec); // comment it - if track_id is not required
+ auto end = std::chrono::steady_clock::now();
+ std::chrono::duration<double> spent = end - start;
+ std::cout << " Time: " << spent.count() << " sec \n";
+
+ //result_vec = detector.tracking_id(result_vec); // comment it - if track_id is not required
draw_boxes(mat_img, result_vec, obj_names);
+ cv::imshow("window name", mat_img);
show_console_result(result_vec, obj_names);
+ cv::waitKey(0);
}
#else
//std::vector<bbox_t> result_vec = detector.detect(filename);
--
Gitblit v1.10.0