deshwalmahesh · April 19, 2021 16:02 · Apr 19, 2021 · Apr 17, 2021 · Apr 17, 2021 · Apr 17, 2021
diff --git a/generate_yolo_annot.py b/generate_yolo_annot.py
@@ -148,3 +148,20 @@ def IoU(true_bb:[tuple,list,np.ndarray]=(0,0,0,0), pred_bb:[tuple,list,np.ndarra
 
     iou = intersect_area/float(union_area)
     return iou
+
+
+def pad_resize(img:np.ndarray,width:int=224,height:int=224)->np.ndarray:
+  '''
+  Pad or Resize the Image for given dimensions. For increasing the size, it keeps the Image in middle by adding padding any color.
+  If one dimension (width or height) has to be increased or other has to be decreased, then it'll pad the short size and then resize the bigger dimension
+  '''
+  x, y, c = img.shape
+  if height > y or width > x:
+    x_ = (width - x)//2 if width > x else 0
+    y_ = (height - y)//2 if height > y else 0
+    img = np.pad(img,((y_,y_),(x_,x_),(0,0)),constant_values=255)
+    x, y, c = img.shape
+  if height < y or width < x:
+    img = cv2.resize(img, dsize=(width, height), interpolation=cv2.INTER_CUBIC)
+
+  return img
diff --git a/generate_yolo_annot.py b/generate_yolo_annot.py
@@ -124,7 +124,7 @@ def generate_text_annotation(dir_path:str,weight_file_path:str,config_file_path:
             f.write('Default Class')
 
 
-def IoU(true_bb:[tuple,list,np.ndarray], pred_bb:[tuple,list,np.ndarray])->float:
+def IoU(true_bb:[tuple,list,np.ndarray]=(0,0,0,0), pred_bb:[tuple,list,np.ndarray]=(0,0,0,0))->float:
     '''
     Get the Intersection Over Union of two Bounding Boxes
     args:

diff --git a/generate_yolo_annot.py b/generate_yolo_annot.py
@@ -122,3 +122,29 @@ def generate_text_annotation(dir_path:str,weight_file_path:str,config_file_path:
 
     with open(dir_path+'classes.txt','w')as f:
             f.write('Default Class')
+
+
+def IoU(true_bb:[tuple,list,np.ndarray], pred_bb:[tuple,list,np.ndarray])->float:
+    '''
+    Get the Intersection Over Union of two Bounding Boxes
+    args:
+        true_bb: Coordinates of True Bounding Box given as (xmin, y_min, x_max, y_max)
+        true_bb: Coordinates of Predicted Bounding Box given as (xmin, y_min, x_max, y_max)
+    out:
+        floating value between 0 and 1 defining IoU of two boxes
+    '''
+    # Open Image ans see numerator blue box to understand the logic for xA, yA, xB, yB https://www.pyimagesearch.com/wp-content/uploads/2016/09/iou_equation.png
+    xA = max(true_bb[0], pred_bb[0]) # xA is the X_min is the max of 2 which will act as the x_min for Intersection Box
+    yA = max(true_bb[1], pred_bb[1]) # Same as xA
+    xB = min(true_bb[2], pred_bb[2]) # xB is the min of two which will act as the x_max for intersection box
+    yB = min(true_bb[3], pred_bb[3]) # same as xB
+
+    intersect_area = max(0, xB - xA+1) * max(0, yB - yA+1) # Width * Height of the blue common box (xB - xA+1) -> Height of blue box. If xa=xb, then Area is 0 that's why max(0,w)
+
+    true_area = (true_bb[2] - true_bb[0] + 1) * (true_bb[3] - true_bb[1] + 1)
+    pred_area = (pred_bb[2] - pred_bb[0] + 1) * (pred_bb[3] - pred_bb[1] + 1)
+
+    union_area = true_area + pred_area - intersect_area # subtract the common area once as it will be included 2 times. One for true area and one for pred area
+
+    iou = intersect_area/float(union_area)
+    return iou
diff --git a/generate_yolo_annot.py b/generate_yolo_annot.py
@@ -41,8 +41,31 @@ def bnd_box_to_yolo_line(box:np.ndarray,img_size:np.ndarray)->tuple:
     h = float((y_max - y_min)) / img_size[0]
 
     return x_center, y_center, w, h
-
-
+
+def yolo_to_bb(annotations:[tuple,list,np.ndarray], img_size:tuple, return_wh:bool = True):
+    '''
+    Change the YoloV4 Darknet annotations to Bounding Box Format
+    args:
+        annotations: Annotations for Darknet YoloV4 format gives as (xc,yc,w,h) iun the .txt file. Exclude class
+        img_size: Size of the original Image in format (w,h,c)
+        return_wh: Whether to return Widrh, Height or the Max_X, Max_Y
+    out:
+        Tuple of Values of a Bounding Box as (x_min, y_min, x_max, y_max) or (x_min, y_min, w, h) depending on the third arg
+    '''
+    (xc,yc,w,h) = annotations
+
+    x_min = (xc*img_size[1]) - ((w*img_size[1])/2)
+    x_max = (xc * img_array.shape[1] *2 ) - x_min
+
+    y_min = (yc*img_size[0]) - ((h*img_size[0])/2)
+    y_max = (yc * img_array.shape[0] *2 ) - y_min
+
+    if return_wh:
+        return int(x_min), int(y_min), int(x_max) - int(x_min) , int(y_max) - int(y_min)  
+
+    return int(x_min), int(y_min), int(x_max), int(y_max)
+
+
 
 def build_yolo_model(weight_file_path:str,config_file_path:str,size:tuple=(416,416)):
     '''

diff --git a/generate_yolo_annot.py b/generate_yolo_annot.py
@@ -0,0 +1,101 @@
+import cv2
+from PIL import Image
+import numpy as np
+import glob
+
+def select_box(results:np.ndarray,method:str)->int:
+    '''
+    Select a Single BB based on Max Probability or Max area logic
+    args:
+        results: Pass in the results by detection module in (classes, scores, boxes) format
+        method: Whether to use 'prob' or 'area'
+    out:
+        Index of the bounding box to select
+    '''
+    classes, scores, bboxes = results
+
+    if method == 'area':
+        return np.argmax([box[2]*box[3] for box in bboxes ])
+
+    return np.argmax(scores)
+
+
+def bnd_box_to_yolo_line(box:np.ndarray,img_size:np.ndarray)->tuple:
+    '''
+    Change Bounding Box to YOLO text format
+    args:
+        box: 1 D array of Bounding box in format [x,y,w,h]
+        img_size: 1-D array of Image Size in format [Width, Height, Channels]
+    out:
+        4 floating point values as (x_center, y_center) are relative points of center of Rectangle. (w,h) is the width and height of Rectangle Relative to Image Size
+    '''
+    (x_min, y_min) = (box[0], box[1])
+    (w, h) = (box[2], box[3])
+    x_max = x_min + w
+    y_max = y_min + h
+
+    x_center = float((x_min + x_max)) / 2 / img_size[1]
+    y_center = float((y_min + y_max)) / 2 / img_size[0]
+
+    w = float((x_max - x_min)) / img_size[1]
+    h = float((y_max - y_min)) / img_size[0]
+
+    return x_center, y_center, w, h
+
+
+
+def build_yolo_model(weight_file_path:str,config_file_path:str,size:tuple=(416,416)):
+    '''
+    Build a Yolo Model
+    args:
+        weight_file_path: Path to the .weights (Yolo v3,v4 etc) file
+        config_file_path: PAth to the .cfg file
+        size: Size of the model detection. You can pass in multiple of 32. Works even when you have trained with 416 and now testing on 608
+    '''
+    net = cv2.dnn.readNet(weight_file_path, config_file_path) 
+    model = cv2.dnn_DetectionModel(net)
+    model.setInputParams(size=size, scale=1/255.)
+    return model
+
+def generate_text_annotation(dir_path:str,weight_file_path:str,config_file_path:str,size:tuple=(416,416),CONFIDENCE_THRESHOLD:float=0.51, NMS_THRESHOLD:float=0.51)->None:
+    '''
+    Generate Annotation File per image for images given in a directory. Uses Bounding Box from the model
+    args:
+        dir_path: Directory path where your images are downloaded. (./dir/whatever/) We are Assuming that they are in .png format only
+        weight_file_path: Path to the .weights (Yolo v3,v4 etc) file
+        config_file_path: PAth to the .cfg file
+        size: Size of the model detection. You can pass in multiple of 32. Works even when you have trained with 416 and now testing on 608
+        CONFIDENCE_THRESHOLD: Only MAke detections valid when the Confidence is above thsi level. Increasing this will lead to FN and decreasing will lead to FP
+        NMS_THRESHOLD: Non MAximum Suppression threshold. Decreasing this will give more number of BB per image. Increasing it will give less no of BBs
+    '''
+    net = cv2.dnn.readNet(weight_file_path, config_file_path) 
+    model = cv2.dnn_DetectionModel(net)
+    model.setInputParams(size=size, scale=1/255.)
+
+
+    image_names = glob.glob(f'{dir_path}*.png')
+    for image_path in image_names:
+        annot = [0] # one class annotation. by default 0
+        text_file_name = '.'+image_path.split('.')[1]+'.txt'
+        img_array =  np.array(Image.open(image_path))
+
+        classes, scores, bboxes = model.detect(img_array, CONFIDENCE_THRESHOLD, NMS_THRESHOLD)
+
+        if type(classes) == tuple:
+            continue
+        if classes.shape == (1,1): # if onlt 1 detection
+            index = 0    
+        else:
+            index = select_box((classes, scores, bboxes),'prob') # if multiple select on Max Area vs Max Prob
+
+
+        box = bboxes[index]
+        score = scores[index]
+
+        annot.extend(bnd_box_to_yolo_line(box,img_array.shape))
+
+        with open(text_file_name,'w')as f:
+            f.write(' '.join([str(i) for i in annot]))
+
+    with open(dir_path+'classes.txt','w')as f:
+            f.write('Default Class')