2017-10-14

springkim · springkim · commit 38584c313f75 · 2017-10-14T14:15:05.000+09:00
diff --git a/CNTK/FasterRCNN/src/.idea/workspace.xml b/CNTK/FasterRCNN/src/.idea/workspace.xml
diff --git a/CNTK/FasterRCNN/src/FasterRCNN_Detect_SE.py b/CNTK/FasterRCNN/src/FasterRCNN_Detect_SE.py
@@ -11,6 +11,8 @@
 import utils.od_utils as od
 from utils.FasterRCNN_eval import FasterRCNN_Evaluator
 
+import re
+
 def get_configuration():
     from utils.configs.FasterRCNN_config import cfg as detector_cfg
     from utils.configs.VGG16_config import cfg as network_cfg
@@ -21,19 +23,36 @@ def prepare_detect(cfg, use_arg_parser=True):
     cfg.MB_SIZE = 1
     cfg.NUM_CHANNELS = 3
     np.random.seed(seed=cfg.RND_SEED)
+
+def imreadEX(image_path):
+    if re.compile('[^ ㄱ-ㅣ가-힣]+').sub('', image_path):
+        stream = open(image_path, "rb")
+        bytes = bytearray(stream.read())
+        numpyarray = np.asarray(bytes, dtype=np.uint8)
+        img = cv2.imdecode(numpyarray, cv2.IMREAD_UNCHANGED)
+    else:
+        img = cv2.imread(image_path)
+    return img
+
 STANDARD_RIGHTS_REQUIRED = 0x000F0000
 SYNCHRONIZE = 0x00100000
 MUTANT_QUERY_STATE = 0x0001
 MUTEX_ALL_ACCESS = STANDARD_RIGHTS_REQUIRED | SYNCHRONIZE | MUTANT_QUERY_STATE
 #
 #   argv[1]=shmem key
 #   argv[2]=mutex key
-#   argv[3]=model path
+#   argv[3]=shmem size
+#   argv[4]=model path
+#   argv[5]=default threshold
 #
 shmem_key=sys.argv[1]
 mutex_key=sys.argv[2]
 shmem_size=int(sys.argv[3])
 model_path=sys.argv[4]
+threshold=float(sys.argv[5])
+
+
+
 shm = mmap.mmap(0,shmem_size,shmem_key)
 mutex = win32event.OpenMutex(MUTEX_ALL_ACCESS, False, mutex_key)
 if not (mutex):
@@ -59,16 +78,17 @@ def prepare_detect(cfg, use_arg_parser=True):
 shm.write(bytes(str(1), "ascii"))
 win32event.ReleaseMutex(mutex)
 #=======================================================================================================================
-while(True):
+while(True ):
     win32event.WaitForSingleObject(mutex,win32event.INFINITE)
     shm.seek(0)
     run_type = shm.read(1)
     if(run_type == b'2'):
-        img_path = str(shm.read(256),"ascii")
-        #detect
+        img_path = shm.read(256).decode('cp949');
+        img_path = img_path.split('\0', 1)[0]
+        # detect
         regressed_rois, cls_probs = od.predict_single_image(model, img_path, cfg, evaluator)
-        #fix roi
-        img = cv2.imread(img_path)
+        # fix roi
+        img = imreadEX(img_path)
         height, width = img.shape[:2]
         scale = 850.0 / max(width, height)
         pad = int((max(height, width) - min(height, width)) / 2)
@@ -81,7 +101,7 @@ def prepare_detect(cfg, use_arg_parser=True):
                 if(m_score<cls_probs[i][j]):
                     m_score=cls_probs[i][j]
                     m_class=j
-            if(m_class!=0 and cls_probs[i][m_class]>0.4):
+            if(m_class!=0 and cls_probs[i][m_class]>threshold):
                 regressed_rois[i] /= scale
                 if (width > height):
                     regressed_rois[i][1] -= pad
diff --git a/CNTK/FasterRCNN/src/FasterRCNN_Detect_Test.py b/CNTK/FasterRCNN/src/FasterRCNN_Detect_Test.py
@@ -0,0 +1,90 @@
+import os
+import sys
+import time
+import numpy as np
+import cv2
+import re
+import mmap
+import win32event
+import win32con
+from cntk import load_model
+from utils.config_helpers import merge_configs
+import utils.od_utils as od
+from utils.FasterRCNN_eval import FasterRCNN_Evaluator
+
+def get_configuration():
+    from utils.configs.FasterRCNN_config import cfg as detector_cfg
+    from utils.configs.VGG16_config import cfg as network_cfg
+    from user_config import cfg as dataset_cfg
+    return merge_configs([detector_cfg, network_cfg, dataset_cfg])
+
+def prepare_detect(cfg, use_arg_parser=True):
+    cfg.MB_SIZE = 1
+    cfg.NUM_CHANNELS = 3
+    np.random.seed(seed=cfg.RND_SEED)
+
+def imreadEX(image_path):
+    if re.compile('[^ ㄱ-ㅣ가-힣]+').sub('', image_path):
+        stream = open(image_path, "rb")
+        bytes = bytearray(stream.read())
+        numpyarray = np.asarray(bytes, dtype=np.uint8)
+        img = cv2.imdecode(numpyarray, cv2.IMREAD_UNCHANGED)
+    else:
+        img = cv2.imread(image_path)
+    return img
+
+STANDARD_RIGHTS_REQUIRED = 0x000F0000
+SYNCHRONIZE = 0x00100000
+MUTANT_QUERY_STATE = 0x0001
+MUTEX_ALL_ACCESS = STANDARD_RIGHTS_REQUIRED | SYNCHRONIZE | MUTANT_QUERY_STATE
+
+cfg=get_configuration()
+prepare_detect(cfg, False)
+
+model=load_model('../faster_rcnn_eval_VGG16_e2e.model')
+# make evaluator
+evaluator = FasterRCNN_Evaluator(model, cfg)
+# prepare main image detect. BCZ It is too slow first,second time.
+od.predict_dummy_image(model,  cfg,evaluator)
+od.predict_dummy_image(model,  cfg,evaluator)
+od.predict_dummy_image(model,  cfg,evaluator)
+od.predict_dummy_image(model,  cfg,evaluator)
+od.predict_dummy_image(model,  cfg,evaluator)
+# model load finished
+
+
+img_path = '../이미지/0000.jpg'
+
+regressed_rois, cls_probs = od.predict_single_image(model, img_path, cfg, evaluator)
+
+img = imreadEX(img_path)
+height, width = img.shape[:2]
+scale = 850.0 / max(width, height)
+pad = int((max(height, width) - min(height, width)) / 2)
+
+for i in range(0, len(regressed_rois)):
+    m_class=0
+    m_score=0
+    for j in range(0,len(cls_probs[i])):
+        if(m_score<cls_probs[i][j]):
+            m_score=cls_probs[i][j]
+            m_class=j
+    if(m_class!=0 and cls_probs[i][m_class]>0.4):
+        regressed_rois[i] /= scale
+        if (width > height):
+            regressed_rois[i][1] -= pad
+            regressed_rois[i][3] -= pad
+        else:
+            regressed_rois[i][0] -= pad
+            regressed_rois[i][2] -= pad
+        cv2.rectangle(img, (regressed_rois[i][0],regressed_rois[i][1]),(regressed_rois[i][2],regressed_rois[i][3]),(0,0,255))
+
+cv2.imshow('result',img)
+cv2.waitKey()
+
+#http://blog.csdn.net/yatere/article/details/7237023
+#https://mail.python.org/pipermail/python-list/2005-April/311564.html
+#https://stackoverflow.com/questions/15526665/wait-until-a-mutex-is-created-in-python-win32
+#https://stackoverflow.com/questions/26114518/ipc-between-python-and-win32-on-windows-os
+#=====EXE=====
+#http://infocentre.tistory.com/3
diff --git a/CNTK/FasterRCNN/src/__pycache__/user_config.cpython-35.pyc b/CNTK/FasterRCNN/src/__pycache__/user_config.cpython-35.pyc
diff --git a/CNTK/FasterRCNN/src/user_config.py b/CNTK/FasterRCNN/src/user_config.py
@@ -13,14 +13,14 @@
 
 # data set config
 __C.DATA.DATASET = "CNTK_FasterRCNN_SpringEdition"
-#__C.DATA.MAP_FILE_PATH = "../../TrainDataSet"
+#__C.DATA.MAP_FILE_PATH = "../../TrainDataSet"          #not used
 __C.DATA.CLASS_MAP_FILE = "class_map.txt"
 __C.DATA.TRAIN_MAP_FILE = "train_img_file.txt"
 __C.DATA.TRAIN_ROI_FILE = "train_roi_file.txt"
-#__C.DATA.TEST_MAP_FILE = "test_img_file.txt"
-#__C.DATA.TEST_ROI_FILE = "test_roi_file.txt"
-#__C.DATA.NUM_TRAIN_IMAGES = 60
-#__C.DATA.NUM_TEST_IMAGES = 10
+#__C.DATA.TEST_MAP_FILE = "test_img_file.txt"          #not used
+#__C.DATA.TEST_ROI_FILE = "test_roi_file.txt"          #not used
+#__C.DATA.NUM_TRAIN_IMAGES = 60                       #automatically select
+#__C.DATA.NUM_TEST_IMAGES = 10                        #not used
 __C.DATA.PROPOSAL_LAYER_SCALES = [8,16,32]
 
 
diff --git a/CNTK/FasterRCNN/src/utils/__pycache__/od_reader.cpython-35.pyc b/CNTK/FasterRCNN/src/utils/__pycache__/od_reader.cpython-35.pyc
diff --git a/CNTK/FasterRCNN/src/utils/__pycache__/plot_helpers.cpython-35.pyc b/CNTK/FasterRCNN/src/utils/__pycache__/plot_helpers.cpython-35.pyc
diff --git a/CNTK/FasterRCNN/src/utils/od_reader.py b/CNTK/FasterRCNN/src/utils/od_reader.py
@@ -9,7 +9,7 @@
 import numpy as np
 import os
 from utils.proposal_helpers import ProposalProvider, compute_targets, compute_image_stats
-
+import re
 DEBUG = False
 if DEBUG:
     import matplotlib.pyplot as mp
@@ -147,7 +147,13 @@ def _read_image(self, image_path):
             imgnp = np.array(bytearray(imgdata), dtype=np.uint8)
             img = cv2.imdecode(imgnp, 1)
         else:
-            img = cv2.imread(image_path)
+            if re.compile('[^ ㄱ-ㅣ가-힣]+').sub('', image_path):
+                stream = open(image_path, "rb")
+                bytes = bytearray(stream.read())
+                numpyarray = np.asarray(bytes, dtype=np.uint8)
+                img = cv2.imdecode(numpyarray, cv2.IMREAD_UNCHANGED)
+            else:
+                img = cv2.imread(image_path)
 
         return img
 
diff --git a/CNTK/FasterRCNN/src/utils/plot_helpers.py b/CNTK/FasterRCNN/src/utils/plot_helpers.py
@@ -17,6 +17,7 @@
 from utils.nms_wrapper import apply_nms_to_single_image_results
 from utils.rpn.bbox_transform import regress_rois
 import cv2 # pip install opencv-python
+import re
 
 available_font = "arial.ttf"
 try:
@@ -31,8 +32,14 @@ def load_resize_and_pad(image_path, width, height, pad_value=114):
     if "@" in image_path:
         print("WARNING: zipped image archives are not supported for visualizing results.")
         exit(0)
-
-    img = cv2.imread(image_path)
+    if re.compile('[^ ㄱ-ㅣ가-힣]+').sub('', image_path):
+        stream = open(image_path, "rb")
+        bytes = bytearray(stream.read())
+        numpyarray = np.asarray(bytes, dtype=np.uint8)
+        img = cv2.imdecode(numpyarray, cv2.IMREAD_UNCHANGED)
+    else:
+        img = cv2.imread(image_path)
+    #img = cv2.imread(image_path)
     return resize_and_pad(img, width, height, pad_value)
 def load_dummy_image_resize_and_pad(width, height, pad_value=114):
     img = np.zeros((64, 64, 3), np.uint8)
diff --git a/FasterRCNN_SE_Detection_Example/.vs/FasterRCNN_SE_Detection_Example/v14/.suo b/FasterRCNN_SE_Detection_Example/.vs/FasterRCNN_SE_Detection_Example/v14/.suo
diff --git a/FasterRCNN_SE_Detection_Example/3rdparty/include/ispring/modules/CV/Eval.h b/FasterRCNN_SE_Detection_Example/3rdparty/include/ispring/modules/CV/Eval.h
@@ -79,7 +79,13 @@ namespace ispring {
 		*	@return pair.first는 Recall 이고 pair.second는 precision 입니다.
 		*/
 		static std::pair<float,float> GetRecallPrecisionSingleClass(std::vector<BoxSE> ground_truth,std::vector<BoxSE> predict,float threshold1=0.5F) {
-			std::pair<float, float> ret;
+			std::pair<float, float> ret = { 0.0F,0.0F };
+			if (predict.size() == 0 && ground_truth.size() == 0) {
+				return std::make_pair(1.0F, 1.0F);
+			}
+			if (predict.size() == 0 || ground_truth.size() == 0) {
+				return std::make_pair(0.0F, 0.0F);
+			}
 			std::vector<int> correct;
 			correct.assign(predict.size(), -1);
 			for (size_t i = 0; i < predict.size(); ++i) {
@@ -187,6 +193,28 @@ namespace ispring {
 			cv::rectangle(img, or +cv::Point(0, -text.height - padding * 2), or +cv::Point(text.width, 0), c, CV_FILLED);
 			cv::putText(img, label, or +cv::Point(0, -padding), font_face, scale, text_color, thickness);
 		}
+		/**
+		*	@brief BoxSE를 이미지에 색 박스로 그립니다.
+		*	@param img 박스가 그려질 이미지
+		*	@param box 박스
+		*	@param c 색상. 지정하지 않으면 GetRGB 함수를 사용합니다.
+		*	@remark
+		*		<img src="https://i.imgur.com/gCWZzVG.jpg" width="640">
+		*/
+		static void DrawColorBoxSE(cv::Mat img, BoxSE rect, cv::Scalar c) {
+			auto MakeColorBox = [](int w, int h, cv::Scalar c)->cv::Mat {
+				std::vector<cv::Mat> planes;
+				planes.push_back(cv::Mat::zeros(h, w, CV_8UC1) + c[2]);
+				planes.push_back(cv::Mat::zeros(h, w, CV_8UC1) + c[1]);
+				planes.push_back(cv::Mat::zeros(h, w, CV_8UC1) + c[0]);
+				cv::Mat box;
+				cv::merge(planes, box);
+				return box;
+			};
+			cv::Mat crop = img(rect);
+			cv::Mat box = MakeColorBox(rect.width, rect.height,c);
+			crop = 0.3*crop + 0.7*box;
+		}
 	};
 }
 #endif  //ISPRING_7E1_A_5_EVAL_HPP_INCLUDED
diff --git a/FasterRCNN_SE_Detection_Example/3rdparty/include/ispring/modules/CV/Image.h b/FasterRCNN_SE_Detection_Example/3rdparty/include/ispring/modules/CV/Image.h
@@ -471,6 +471,7 @@ namespace ispring {
 			}
 			return glue;
 		}
+		
 #if defined(ISPRING_WINDOWS) || defined(DOXYGEN)
 		/**
 		*	@brief HBITMAP을 cv::Mat 으로 변환 합니다.
diff --git a/FasterRCNN_SE_Detection_Example/3rdparty/include/ispring/modules/Console/xout.h b/FasterRCNN_SE_Detection_Example/3rdparty/include/ispring/modules/Console/xout.h
@@ -44,20 +44,20 @@ auto get_endl(const std::basic_ostream<e, t>&)-> decltype(&std::endl<e, t>) {
 #if _MSC_VER==1910		///VS 2017
 //https://i.imgur.com/TOFHpX4.png
 #if _WIN64
-const char* compiler = "C:/Program Files (x86)/Microsoft Visual Studio/2017/Community/VC/Auxiliary/Build/vcvarsall.bat amd64 ";
+SELECT_ANY const char* compiler = "C:/Program Files (x86)/Microsoft Visual Studio/2017/Community/VC/Auxiliary/Build/vcvarsall.bat amd64 ";
 #else
-const char* compiler = "C:/Program Files (x86)/Microsoft Visual Studio/2017/Community/VC/Auxiliary/Build/vcvarsall.bat x86 ";
+SELECT_ANY const char* compiler = "C:/Program Files (x86)/Microsoft Visual Studio/2017/Community/VC/Auxiliary/Build/vcvarsall.bat x86 ";
 #endif
 #elif _MSC_VER==1900	///VS 2015
-const char* compiler = "C:/Program Files (x86)/Microsoft Visual Studio 14.0/VC/vcvarsall.bat";
+SELECT_ANY const char* compiler = "C:/Program Files (x86)/Microsoft Visual Studio 14.0/VC/vcvarsall.bat";
 #elif _MSC_VER==1800		///VS 2013
-const char* compiler = "C:/Program Files (x86)/Microsoft Visual Studio 12.0/VC/vcvarsall.bat";
+SELECT_ANY const char* compiler = "C:/Program Files (x86)/Microsoft Visual Studio 12.0/VC/vcvarsall.bat";
 #elif _MSC_VER==1700		///VS2012
-const char* compiler = "C:/Program Files (x86)/Microsoft Visual Studio 11.0/VC/vcvarsall.bat";
+SELECT_ANY const char* compiler = "C:/Program Files (x86)/Microsoft Visual Studio 11.0/VC/vcvarsall.bat";
 #elif _MSC_VER==1600		///VS2010
-const char* compiler = "C:/Program Files (x86)/Microsoft Visual Studio 10.0/VC/vcvarsall.bat";
+SELECT_ANY const char* compiler = "C:/Program Files (x86)/Microsoft Visual Studio 10.0/VC/vcvarsall.bat";
 #elif defined(__GNUC__) ///MinGW
-const char* compiler = "g++";
+SELECT_ANY const char* compiler = "g++";
 #endif
 #endif //DOXYGEN
 namespace ispring { 
diff --git a/FasterRCNN_SE_Detection_Example/3rdparty/include/ispring/version_info.txt b/FasterRCNN_SE_Detection_Example/3rdparty/include/ispring/version_info.txt
@@ -1 +1 @@
-<ispring-version>1.1.3</ispring-version>
+<ispring-version>1.1.6</ispring-version>
diff --git a/FasterRCNN_SE_Detection_Example/FasterRCNN_SE.h b/FasterRCNN_SE_Detection_Example/FasterRCNN_SE.h
@@ -60,7 +60,7 @@ class FasterRCNN {
 	//9(57) = terminate signal
 	std::vector<std::string> m_class_map;
 public:
-	void Create(std::string model_path,std::string classfile,DWORD size=6000) {
+	void Create(std::string model_path,std::string classfile,DWORD size=10000,float filter_threshold=0.4F) {
 		m_key_shmem = this->GetKey() + "_shmem";
 		m_key_mutex = this->GetKey() + "_mutex";
 		m_size = size;
@@ -72,7 +72,7 @@ class FasterRCNN {
 		//std::string exec = "python \"C:/Users/spring/Documents/SourceTree/FasterRCNN_SpringEdition/CNTK/FasterRCNN/src/FasterRCNN_Detect_SE.py\"";
 		std::string exec = "FasterRCNN_Detect_SE.exe";
 		std::ostringstream oss;
-		oss << exec << " " << m_key_shmem << " " << m_key_mutex << " " << m_size << " " << "\"" << model_path << "\"";
+		oss << exec << " " << m_key_shmem << " " << m_key_mutex << " " << m_size << " " << "\"" << model_path << "\"" << "\t" << filter_threshold;
 		UINT ret=WinExec(oss.str().c_str(), SW_SHOW);
 		if (ret < 31) {
 			::MessageBoxA(nullptr, "FasterRCNN Detector execute failed\nIt needs \"FasterRCNN_Detect_SE.exe\".", "Error", MB_OK);
@@ -132,13 +132,13 @@ class FasterRCNN {
 		while (pass == false) {
 			WaitForSingleObject(m_mutex, INFINITE);
 			if (m_buffer[0] == '1') {
-				
 				memset(m_buffer + 1, 0,m_size - 1);
 				char* p = m_buffer + 1;
 				const char* q = img_path.data();
 				while (*q != '\0') {
 					*p++ = *q++;
 				}
+				*p = '\0';
 				//strcpy(m_buffer + 1, img_path.c_str());	//strcpy is not working on SDL.
 				//strcpy_s(m_buffer + 1, m_size-1,img_path.c_str());		//strcpy_s is not working on debug mode.
 				m_buffer[0] = '2';	//running
diff --git a/FasterRCNN_SE_Detection_Example/main.cpp b/FasterRCNN_SE_Detection_Example/main.cpp
@@ -1,5 +1,4 @@
-﻿// dllmain.cpp: DLL 응용 프로그램의 진입점을 정의합니다.
-#include <Windows.h>
+﻿#include <Windows.h>
 #include<string>
 #include<iostream>
 #include"ispring/All.h"	//@see https://github.com/springkim/ISpring
diff --git a/README.md b/README.md
@@ -6,7 +6,8 @@
 * You need only 1 header file, To detect by FasterRCNN.
 
 <img src="https://i.imgur.com/ElCyyzT.png" title="Windows8" width="48"><img src="https://i.imgur.com/O5bye0l.png" width="48">
-
+## Lastest news
+**2017-10-14.** : Support unicode(korean) path. Changed default IPC buffers as 10000 bytes.
 ## Setup for train
 #### 1. Download model and files.
 There is a **download_model.bat** in [CNTK/FasterRCNN/](https://github.com/springkim/FasterRCNN_SpringEdition/tree/master/CNTK/FasterRCNN).
