add YOLOv3 support

extra/README.md

@@ -31,10 +31,10 @@
 
 - ### convert YOLO to our format:
 
-    1) Add class list to the file `class_list.txt`
-    2) Insert predicted objects files into **predicted/**
-    3) Insert images into **images/**
-    4) Run the python script: `python convert_pred_yolo.py`
+    After runnuning darknet on a list of images, e.g.: `darknet.exe detector test data/voc.data yolo-voc.cfg yolo-voc.weights -dont_show -ext_output < data/train.txt > result.txt`
+
+    1) Copy the file `result.txt` to the folder `extra/`
+    2) Run the python script: `python convert_pred_yolo.py`
 
 - ### convert keras-yolo3 to our format:
 

extra/convert_pred_yolo.py

@@ -1,83 +1,36 @@
-import sys
-import os
-import glob
-import cv2
-
-
-def convert_yolo_coordinates_to_voc(x_c_n, y_c_n, width_n, height_n, img_width, img_height):
-  ## remove normalization given the size of the image
-  x_c = float(x_c_n) * img_width
-  y_c = float(y_c_n) * img_height
-  width = float(width_n) * img_width
-  height = float(height_n) * img_height
-  ## compute half width and half height
-  half_width = width / 2
-  half_height = height / 2
-  ## compute left, top, right, bottom
-  ## in the official VOC challenge the top-left pixel in the image has coordinates (1;1)
-  left = int(x_c - half_width) + 1
-  top = int(y_c - half_height) + 1
-  right = int(x_c + half_width) + 1
-  bottom = int(y_c + half_height) + 1
-  return left, top, right, bottom
-
-# read the class_list.txt to a list
-with open("class_list.txt") as f:
-  obj_list = f.readlines()
-## remove whitespace characters like `\n` at the end of each line
-  obj_list = [x.strip() for x in obj_list]
-## e.g. first object in the list
-#print(obj_list[0])
-
-# change directory to the one with the files to be changed
-path_to_folder = '../predicted'
-#print(path_to_folder)
-os.chdir(path_to_folder)
-
-# old files (YOLO format) will be moved to a new folder (backup/)
-## create the backup dir if it doesn't exist already
-if not os.path.exists("backup"):
-  os.makedirs("backup")
-
-# create VOC format files
-txt_list = glob.glob('*.txt')
-if len(txt_list) == 0:
-  print("Error: no .txt files found in predicted")
-  sys.exit()
-for tmp_file in txt_list:
-  #print(tmp_file)
-  # 1. check that there is an image with that name
-  ## get name before ".txt"
-  image_name = tmp_file.split(".txt",1)[0]
-  #print(image_name)
-  ## check if image exists
-  for fname in os.listdir('../images'):
-    if fname.startswith(image_name):
-      ## image found
-      #print(fname)
-      img = cv2.imread('../images/' + fname)
-      ## get image width and height
-      img_height, img_width = img.shape[:2]
-      break
-  else:
-    ## image not found
-    print("Error: image not found, corresponding to " + tmp_file)
-    sys.exit()
-  # 2. open txt file lines to a list
-  with open(tmp_file) as f:
-    content = f.readlines()
-  ## remove whitespace characters like `\n` at the end of each line
-  content = [x.strip() for x in content]
-  # 3. move old file (YOLO format) to backup
-  os.rename(tmp_file, "backup/" + tmp_file)
-  # 4. create new file (VOC format)
-  with open(tmp_file, "a") as new_f:
-    for line in content:
-      ## split a line by spaces.
-      ## "c" stands for center and "n" stands for normalized
-      obj_id, conf, x_c_n, y_c_n, width_n, height_n = line.split()
-      obj_name = obj_list[int(obj_id)]
-      left, top, right, bottom = convert_yolo_coordinates_to_voc(x_c_n, y_c_n, width_n, height_n, img_width, img_height)
-      #print(obj_name + " " + str(conf) + " " + str(left) + " " + str(top) + " " + str(right) + " " + str(bottom))
-      new_f.write(obj_name + " " + str(conf) + " " + str(left) + " " + str(top) + " " + str(right) + " " + str(bottom) + '\n')
-print("Conversion completed!")
+import os
+import re
+
+IN_FILE = 'result.txt'
+OUTPUT_DIR = os.path.join('..', 'predicted')
+
+SEPARATOR_KEY = 'Enter Image Path:'
+IMG_FORMAT = '.jpg'
+
+outfile = None
+with open(IN_FILE) as infile:
+    for line in infile:
+        if SEPARATOR_KEY in line:
+            if IMG_FORMAT not in line:
+                break
+            # get text between two substrings (SEPARATOR_KEY and IMG_FORMAT)
+            image_path = re.search(SEPARATOR_KEY + '(.*)' + IMG_FORMAT, line)
+            # get the image name (the final component of a image_path)
+            # e.g., from 'data/horses_1' to 'horses_1'
+            image_name = os.path.basename(image_path.group(1))
+            # close the previous file
+            if outfile is not None:
+                outfile.close()
+            # open a new file
+            outfile = open(os.path.join(OUTPUT_DIR, image_name + '.txt'), 'w')
+        elif outfile is not None:
+            # split line on first occurrence of the character ':' and '%'
+            class_name, info = line.split(':', 1)
+            confidence, bbox = info.split('%', 1)
+            # get all the coordinates of the bounding box
+            bbox = bbox.replace(')','') # remove the character ')'
+            # go through each of the parts of the string and check if it is a digit
+            left, top, width, height = [int(s) for s in bbox.split() if s.isdigit()]
+            right = left + width
+            bottom = top + height
+            outfile.write("{} {} {} {} {} {}\n".format(class_name, float(confidence)/100, left, top, right, bottom))

extra/result.txt

@@ -0,0 +1,13 @@
+Total BFLOPS 65.864 
+
+ seen 64 
+Enter Image Path: data/horses.jpg: Predicted in 42.076185 seconds.
+horse: 88%	(left_x:    3   top_y:  185   width:  150   height:  167)
+horse: 99%	(left_x:    5   top_y:  198   width:  307   height:  214)
+horse: 96%	(left_x:  236   top_y:  180   width:  215   height:  169)
+horse: 99%	(left_x:  440   top_y:  209   width:  156   height:  142)
+Enter Image Path: data/person.jpg: Predicted in 41.767213 seconds.
+dog: 99%	(left_x:   58   top_y:  262   width:  147   height:   89)
+person: 100%	(left_x:  190   top_y:   95   width:   86   height:  284)
+horse: 100%	(left_x:  394   top_y:  137   width:  215   height:  206)
+Enter Image Path: