predict.py 文件源码

def get_boxes(ans, block_x, block_y, bb_num,class_num,th, im_w,im_h,biases):
  sorted_boxes = []
  for by in range(block_y):
    for bx in range(block_x):
      for j in range(bb_num):

        box   = ans[by,bx,j,0:4]
        conf  = sigmoid(ans[by,bx,j,4])
        probs = softmax(ans[by,bx,j,5:(5+class_num)])[0]

        p_class = probs*conf

        if np.max(p_class)<th:
          continue
        class_id = np.argmax(p_class)

        x = (bx+sigmoid(box[0]))*(im_w/float(block_x))
        y = (by+sigmoid(box[1]))*(im_h/float(block_y))
        w = np.exp(box[2])*biases[j][0]*(im_w/float(block_x))
        h = np.exp(box[3])*biases[j][1]*(im_h/float(block_y))
        b = Box(x,y,w,h)

        sorted_boxes.append([b,j,class_id, max(p_class)])
  return sorted_boxes