如何在tensorflow对象检测api中绘制边界框而不使用归一化坐标

Question

如何在tensorflow对象检测API中绘制边界框而不使用归一化坐标？在object_detection_tutorial.ipynb中，我注意到默认坐标是normalize坐标，框的形式是[xmin,ymin,xmax,ymax]，如何将它们转换为[image_lengthxmin,image_widthymin,image_lengthxmax,image_widthymax]？我试着用

        boxes[0]=boxes[0]*200
        boxes[1]=boxes[1]*100
        boxes[2]=boxes[2]*200
        boxes[3]=boxes[3]*100

但是错误发生了：

---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
<ipython-input-72-efcec9615ee3> in <module>()
     30                   feed_dict={image_tensor: image_np_expanded})
     31                 boxes[0]=boxes[0]*200
---> 32                 boxes[1]=boxes[1]*100
     33                 boxes[2]=boxes[2]*200
     34                 boxes[3]=boxes[3]*100
IndexError: index 1 is out of bounds for axis 0 with size 1

Answer 1

如果查看research/object_detection/utils/visualization_utils.py，则框[0]是ymin而不是xmin当将这些坐标乘以100或200时，请确保它仍然在图像边界(im_width,im_height）中。您可以尝试框[0]*100，框[1]*-200，框[2]*-100，框[3]*200，这与下面的代码类似。

ymin = boxes[0]*100
xmin = boxes[1]*-200
ymax = boxes[2]*-100
xmax = boxes[3]*200

draw = ImageDraw.Draw(image)
im_width, im_height = image.size
(left, right, top, bottom) = (xmin * im_width, xmax * im_width,
                              ymin * im_height, ymax * im_height)

draw.line([(left, top), (left, bottom), (right, bottom),
               (right, top), (left, top)], width=thickness, fill=color)