face_detection/decode_yolo_v2.py

import numpy as np
import matplotlib.pyplot as plt
from matplotlib.patches import Rectangle
import copy
import tensorflow.keras.backend as K
import cv2
from face_detection.config import cell_size,idx_to_class,class_to_idx,class_colors



def get_objects(y_pred,p=0.5,decode_preds=True):
  global tf_anchors
  output_size=y_pred.shape[1]
  image_size=cell_size*output_size
  
  y_pred=copy.deepcopy(y_pred)
  if decode_preds:
    y_pred[...,0]=K.sigmoid(y_pred[...,0])
    y_pred[...,3:5]=np.clip((K.exp(y_pred[...,3:5])*tf_anchors).numpy(),0,output_size)
  # y_pred[...,3:5]=np.clip(y_pred[...,3:5],0,output_size)
  objs_found=[]
  idxs=np.where(y_pred[...,0]>=p)
  if np.size(idxs):
    for i,obj in enumerate(y_pred[idxs[0],idxs[1],idxs[2],:]):
      # obj (p,x,y,w,h,c_1,c_2,c_3,c_4,c_5.......c_n)
      if decode_preds:
        obj[1:3]=K.sigmoid(obj[1:3]) # x,y
    
      prob=obj[0]
      obj=obj[1:]
      
      obj[4]=np.argmax(obj[4:])
      obj=obj[:5]      

      obj[0]=idxs[1][i]+obj[0]  # center x
      obj[1]=idxs[0][i]+obj[1]  # center y

      obj[0]=np.clip(obj[0]-(obj[2]/2),0,output_size)  # xmin
      obj[1]=np.clip(obj[1]-(obj[3]/2),0,output_size)  # ymin
      
      
      obj_name=idx_to_class[obj[4]]
      
      
      # obj_details={'p':prob,'xywh':list(obj[:-1]/output_size),'class_idx':int(obj[4]),'class':obj_name}  # xywh are scaled 0 to 1
      obj_details=[prob,obj[4],*list(obj[:-1]/output_size)]  # xywh are scaled 0 to 1 [P,C_IDX,X,Y,W,H]
      objs_found.append(obj_details)
  return np.array(objs_found)

def list_get_iou(bboxes1, bboxes2):
  # bboxes has xywh => xmin,ymin,width,height
  bboxes1 = [bboxes1[0],bboxes1[1],bboxes1[0]+bboxes1[2],bboxes1[1]+bboxes1[3]]
  bboxes2 = [bboxes2[0],bboxes2[1],bboxes2[0]+bboxes2[2],bboxes2[1]+bboxes2[3]]

  xA = max(bboxes1[0], bboxes2[0])
  yA = max(bboxes1[1], bboxes2[1])
  xB = min(bboxes1[2], bboxes2[2])
  yB = min(bboxes1[3], bboxes2[3])

  intersection_area = max(0, xB - xA ) * max(0, yB - yA )

  box1_area = (bboxes1[2] - bboxes1[0] ) * (bboxes1[3] - bboxes1[1] )
  box2_area = (bboxes2[2] - bboxes2[0] ) * (bboxes2[3] - bboxes2[1] )

  iou = intersection_area / float(box1_area + box2_area - intersection_area+1e-6)

  return iou

def np_iou(bboxes1,bboxes2):
  # bboxes has xywh => xmin,ymin,width,height
  
  boxes1_x1 = bboxes1[:,0]
  boxes1_y1 = bboxes1[:,1]
  boxes1_x2 = boxes1_x1 + bboxes1[:,2]
  boxes1_y2 = boxes1_y1 + bboxes1[:,3]
  
  boxes2_x1 = bboxes2[:,0]
  boxes2_y1 = bboxes2[:,1]
  boxes2_x2 = boxes2_x1 + bboxes2[:,2]
  boxes2_y2 = boxes2_y1 + bboxes2[:,3]          

  
  xmins = np.maximum(boxes1_x1,boxes2_x1)
  ymins = np.maximum(boxes1_y1,boxes2_y1)
  
  xmaxs = np.minimum(boxes1_x2,boxes2_x2)
  ymaxs = np.minimum(boxes1_y2,boxes2_y2)



  intersection = np.clip((xmaxs-xmins),0,None)*np.clip((ymaxs-ymins),0,None)
  
  union = (boxes1_x2-boxes1_x1)*(boxes1_y2-boxes1_y1) + (boxes2_x2-boxes2_x1)*(boxes2_y2-boxes2_y1)
  ious=intersection/((union-intersection)+1e-6)
  
  return ious

def nms(objs_found,iou_threshold=0.2):
  '''objs_found list of list:[
                                [p,c_idx,x,y,w,h],
                                [p,c_idx,x,y,w,h]
                             ]
  '''
  if objs_found.size<2 or iou_threshold==1: return objs_found

  objs_found=objs_found[np.argsort(objs_found[:,0])[::-1] ]# This was very important
  
  best_boxes=[]
  while len(objs_found)>0:
    obj=objs_found[0]
    best_boxes.append(list(obj))
    objs_found=objs_found[1:].reshape(-1,6)
    
    if len(objs_found)>0:

      same_class_idxs=np.where(objs_found[:,1]==obj[1])[0]  # same class_idx
      same_class_objs=objs_found[same_class_idxs].reshape(-1,6)
      
      ious=np_iou(obj[None,2:],same_class_objs[:,2:])
      
      delete_idxs=same_class_idxs[np.where(ious>= iou_threshold)[0]]

      objs_found=np.delete(objs_found,delete_idxs,axis=0)
    
    
  return best_boxes

def show_objects(img,objs_found,return_img=False):
  plt.imshow(img)
  for i in range(len(objs_found)):  
    p=objs_found[i]['p']
    obj=objs_found[i]['xywh']
    obj_name=objs_found[i]['class']
    plt.gca().add_patch(Rectangle((obj[0],obj[1]),(obj[2]),(obj[3]),linewidth=4,edgecolor=class_colors[obj_name],facecolor='none'))
    plt.text(obj[0],obj[1],obj_name)



def pred_image(img,objs_found,font_scale=2,thickness=4):
  def rescale(obj_found,w,h):
    # xywh
    obj_found[0]*=w
    obj_found[1]*=h
    obj_found[2]*=w
    obj_found[3]*=h
    return obj_found

  for i in range(len(objs_found)):  
    # p,c_idx,x,y,w,h
    p=objs_found[i][0]
    obj_name=objs_found[i][1]
    obj=rescale(objs_found[i][2:],img.shape[1],img.shape[0])
    
    img=cv2.rectangle(img,(int(obj[0]),int(obj[1])),(int(obj[0]+obj[2]),int(obj[1]+obj[3])),(class_colors[obj_name]*255),thickness)
    img=cv2.putText(img,obj_name,(int(obj[0]),int(obj[1])),cv2.FONT_HERSHEY_SIMPLEX,font_scale, (0,0,0), thickness, lineType=cv2.LINE_AA)
    # draw_text(img, "world", font_scale=4, pos=(10, 20 + h), text_color_bg=(255, 0, 0))
  return img