From 9ac78d8b84f6a059d2cefe22a10aa60de5b3feaf Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Thu, 04 Jan 2018 21:58:52 +0000
Subject: [PATCH] Fine tuning, use stopbackward=1 in the cfg-file in that layer where Backward should be stopped.
---
src/yolo_v2_class.cpp | 191 +++++++++++++++++++++++++++++++++++++++++------
1 files changed, 165 insertions(+), 26 deletions(-)
diff --git a/src/yolo_v2_class.cpp b/src/yolo_v2_class.cpp
index 69cd5fc..1777aa2 100644
--- a/src/yolo_v2_class.cpp
+++ b/src/yolo_v2_class.cpp
@@ -1,6 +1,5 @@
#include "yolo_v2_class.hpp"
-
#include "network.h"
extern "C" {
@@ -12,43 +11,42 @@
#include "box.h"
#include "image.h"
#include "demo.h"
-
#include "option_list.h"
-
+#include "stb_image.h"
}
//#include <sys/time.h>
#include <vector>
#include <iostream>
-
+#include <algorithm>
#define FRAMES 3
-#define ROI_PER_DETECTOR 100
-
struct detector_gpu_t{
float **probs;
box *boxes;
network net;
- //image det;
- //image det_s;
image images[FRAMES];
float *avg;
float *predictions[FRAMES];
+ int demo_index;
+ unsigned int *track_id;
};
-
YOLODLL_API Detector::Detector(std::string cfg_filename, std::string weight_filename, int gpu_id)
{
int old_gpu_index;
+#ifdef GPU
cudaGetDevice(&old_gpu_index);
+#endif
detector_gpu_ptr = std::make_shared<detector_gpu_t>();
-
detector_gpu_t &detector_gpu = *reinterpret_cast<detector_gpu_t *>(detector_gpu_ptr.get());
+#ifdef GPU
cudaSetDevice(gpu_id);
+#endif
network &net = detector_gpu.net;
net.gpu_index = gpu_id;
//gpu_index = i;
@@ -56,7 +54,7 @@
char *cfgfile = const_cast<char *>(cfg_filename.data());
char *weightfile = const_cast<char *>(weight_filename.data());
- net = parse_network_cfg(cfgfile);
+ net = parse_network_cfg_custom(cfgfile, 1);
if (weightfile) {
load_weights(&net, weightfile);
}
@@ -74,26 +72,86 @@
detector_gpu.probs = (float **)calloc(l.w*l.h*l.n, sizeof(float *));
for (j = 0; j < l.w*l.h*l.n; ++j) detector_gpu.probs[j] = (float *)calloc(l.classes, sizeof(float));
+ detector_gpu.track_id = (unsigned int *)calloc(l.classes, sizeof(unsigned int));
+ for (j = 0; j < l.classes; ++j) detector_gpu.track_id[j] = 1;
+
+#ifdef GPU
cudaSetDevice(old_gpu_index);
+#endif
}
+
YOLODLL_API Detector::~Detector()
{
detector_gpu_t &detector_gpu = *reinterpret_cast<detector_gpu_t *>(detector_gpu_ptr.get());
layer l = detector_gpu.net.layers[detector_gpu.net.n - 1];
- free(detector_gpu.boxes);
+ free(detector_gpu.track_id);
+
free(detector_gpu.avg);
- free(detector_gpu.predictions);
+ for (int j = 0; j < FRAMES; ++j) free(detector_gpu.predictions[j]);
+ for (int j = 0; j < FRAMES; ++j) if(detector_gpu.images[j].data) free(detector_gpu.images[j].data);
+
for (int j = 0; j < l.w*l.h*l.n; ++j) free(detector_gpu.probs[j]);
+ free(detector_gpu.boxes);
free(detector_gpu.probs);
+
+ int old_gpu_index;
+#ifdef GPU
+ cudaGetDevice(&old_gpu_index);
+ cudaSetDevice(detector_gpu.net.gpu_index);
+#endif
+
+ free_network(detector_gpu.net);
+
+#ifdef GPU
+ cudaSetDevice(old_gpu_index);
+#endif
+}
+
+YOLODLL_API int Detector::get_net_width() const {
+ detector_gpu_t &detector_gpu = *reinterpret_cast<detector_gpu_t *>(detector_gpu_ptr.get());
+ return detector_gpu.net.w;
+}
+YOLODLL_API int Detector::get_net_height() const {
+ detector_gpu_t &detector_gpu = *reinterpret_cast<detector_gpu_t *>(detector_gpu_ptr.get());
+ return detector_gpu.net.h;
}
-YOLODLL_API std::vector<bbox_t> Detector::detect(std::string image_filename, float thresh)
+YOLODLL_API std::vector<bbox_t> Detector::detect(std::string image_filename, float thresh, bool use_mean)
+{
+ std::shared_ptr<image_t> image_ptr(new image_t, [](image_t *img) { if (img->data) free(img->data); delete img; });
+ *image_ptr = load_image(image_filename);
+ return detect(*image_ptr, thresh, use_mean);
+}
+
+static image load_image_stb(char *filename, int channels)
+{
+ int w, h, c;
+ unsigned char *data = stbi_load(filename, &w, &h, &c, channels);
+ if (!data)
+ throw std::runtime_error("file not found");
+ if (channels) c = channels;
+ int i, j, k;
+ image im = make_image(w, h, c);
+ for (k = 0; k < c; ++k) {
+ for (j = 0; j < h; ++j) {
+ for (i = 0; i < w; ++i) {
+ int dst_index = i + w*j + w*h*k;
+ int src_index = k + c*i + c*w*j;
+ im.data[dst_index] = (float)data[src_index] / 255.;
+ }
+ }
+ }
+ free(data);
+ return im;
+}
+
+YOLODLL_API image_t Detector::load_image(std::string image_filename)
{
char *input = const_cast<char *>(image_filename.data());
- image im = load_image_color(input, 0, 0);
+ image im = load_image_stb(input, 3);
image_t img;
img.c = im.c;
@@ -101,21 +159,30 @@
img.h = im.h;
img.w = im.w;
- return detect(img, thresh);
+ return img;
}
-YOLODLL_API std::vector<bbox_t> Detector::detect(image_t img, float thresh)
+YOLODLL_API void Detector::free_image(image_t m)
+{
+ if (m.data) {
+ free(m.data);
+ }
+}
+
+YOLODLL_API std::vector<bbox_t> Detector::detect(image_t img, float thresh, bool use_mean)
{
detector_gpu_t &detector_gpu = *reinterpret_cast<detector_gpu_t *>(detector_gpu_ptr.get());
network &net = detector_gpu.net;
int old_gpu_index;
+#ifdef GPU
cudaGetDevice(&old_gpu_index);
cudaSetDevice(net.gpu_index);
+#endif
//std::cout << "net.gpu_index = " << net.gpu_index << std::endl;
- float nms = .4;
+ //float nms = .4;
image im;
im.c = img.c;
@@ -123,16 +190,27 @@
im.h = img.h;
im.w = img.w;
- image sized = resize_image(im, net.w, net.h);
- layer l = net.layers[net.n - 1];
+ image sized;
+
+ if (net.w == im.w && net.h == im.h) {
+ sized = make_image(im.w, im.h, im.c);
+ memcpy(sized.data, im.data, im.w*im.h*im.c * sizeof(float));
+ }
+ else
+ sized = resize_image(im, net.w, net.h);
- //box *boxes = (box *)calloc(l.w*l.h*l.n, sizeof(box));
- //float **probs = (float **)calloc(l.w*l.h*l.n, sizeof(float *));
- // (int j = 0; j < l.w*l.h*l.n; ++j) probs[j] = (float *)calloc(l.classes, sizeof(float *));
+ layer l = net.layers[net.n - 1];
float *X = sized.data;
- network_predict(net, X);
+ float *prediction = network_predict(net, X);
+
+ if (use_mean) {
+ memcpy(detector_gpu.predictions[detector_gpu.demo_index], prediction, l.outputs * sizeof(float));
+ mean_arrays(detector_gpu.predictions, FRAMES, l.outputs, detector_gpu.avg);
+ l.output = detector_gpu.avg;
+ detector_gpu.demo_index = (detector_gpu.demo_index + 1) % FRAMES;
+ }
get_region_boxes(l, 1, 1, thresh, detector_gpu.probs, detector_gpu.boxes, 0, 0);
if (nms) do_nms_sort(detector_gpu.boxes, detector_gpu.probs, l.w*l.h*l.n, l.classes, nms);
@@ -148,18 +226,79 @@
if (prob > thresh)
{
bbox_t bbox;
- bbox.x = (b.x - b.w / 2.)*im.w;
- bbox.y = (b.y - b.h / 2.)*im.h;
+ bbox.x = std::max((double)0, (b.x - b.w / 2.)*im.w);
+ bbox.y = std::max((double)0, (b.y - b.h / 2.)*im.h);
bbox.w = b.w*im.w;
bbox.h = b.h*im.h;
bbox.obj_id = obj_id;
bbox.prob = prob;
+ bbox.track_id = 0;
bbox_vec.push_back(bbox);
}
}
+ if(sized.data)
+ free(sized.data);
+
+#ifdef GPU
cudaSetDevice(old_gpu_index);
+#endif
return bbox_vec;
+}
+
+YOLODLL_API std::vector<bbox_t> Detector::tracking(std::vector<bbox_t> cur_bbox_vec, int const frames_story)
+{
+ detector_gpu_t &det_gpu = *reinterpret_cast<detector_gpu_t *>(detector_gpu_ptr.get());
+
+ bool prev_track_id_present = false;
+ for (auto &i : prev_bbox_vec_deque)
+ if (i.size() > 0) prev_track_id_present = true;
+
+ if (!prev_track_id_present) {
+ for (size_t i = 0; i < cur_bbox_vec.size(); ++i)
+ cur_bbox_vec[i].track_id = det_gpu.track_id[cur_bbox_vec[i].obj_id]++;
+ prev_bbox_vec_deque.push_front(cur_bbox_vec);
+ if (prev_bbox_vec_deque.size() > frames_story) prev_bbox_vec_deque.pop_back();
+ return cur_bbox_vec;
+ }
+
+ std::vector<unsigned int> dist_vec(cur_bbox_vec.size(), std::numeric_limits<unsigned int>::max());
+
+ for (auto &prev_bbox_vec : prev_bbox_vec_deque) {
+ for (auto &i : prev_bbox_vec) {
+ int cur_index = -1;
+ for (size_t m = 0; m < cur_bbox_vec.size(); ++m) {
+ bbox_t const& k = cur_bbox_vec[m];
+ if (i.obj_id == k.obj_id) {
+ float center_x_diff = (float)(i.x + i.w/2) - (float)(k.x + k.w/2);
+ float center_y_diff = (float)(i.y + i.h/2) - (float)(k.y + k.h/2);
+ unsigned int cur_dist = sqrt(center_x_diff*center_x_diff + center_y_diff*center_y_diff);
+ if (cur_dist < 100 && (k.track_id == 0 || dist_vec[m] > cur_dist)) {
+ dist_vec[m] = cur_dist;
+ cur_index = m;
+ }
+ }
+ }
+
+ bool track_id_absent = !std::any_of(cur_bbox_vec.begin(), cur_bbox_vec.end(),
+ [&i](bbox_t const& b) { return b.track_id == i.track_id && b.obj_id == i.obj_id; });
+
+ if (cur_index >= 0 && track_id_absent){
+ cur_bbox_vec[cur_index].track_id = i.track_id;
+ cur_bbox_vec[cur_index].w = (cur_bbox_vec[cur_index].w + i.w) / 2;
+ cur_bbox_vec[cur_index].h = (cur_bbox_vec[cur_index].h + i.h) / 2;
+ }
+ }
+ }
+
+ for (size_t i = 0; i < cur_bbox_vec.size(); ++i)
+ if (cur_bbox_vec[i].track_id == 0)
+ cur_bbox_vec[i].track_id = det_gpu.track_id[cur_bbox_vec[i].obj_id]++;
+
+ prev_bbox_vec_deque.push_front(cur_bbox_vec);
+ if (prev_bbox_vec_deque.size() > frames_story) prev_bbox_vec_deque.pop_back();
+
+ return cur_bbox_vec;
}
\ No newline at end of file
--
Gitblit v1.10.0