Tensorflow object 检测 api 得到按边界框坐标排序的预测

Question

I trained a neural network to solve simple captcha using tensorflow object-detection API, but when I output the predictions with the following code:

for index, value in enumerate(classes[0]):
object_dict = {}
if scores[0, index] > threshold:
    object_dict[(category_index.get(value)).get('name').encode('utf8')] = scores[0, index]
    objects.append(object_dict)

I get predictions in random order with every function run. I asked a question earlier, and I was advised to try using the coordinates, but I could not find a way to connect the classes and the coordinates of the box that is associated with this class. Example of solved captcha is attached, so I need a way to output predictions in order from left to right.

Image

Answer 1

Given that boxes[0] is an array of shape num_boxes * 4, where the first value in the box is xmin, this can get you the indicies of the boxes, sorted by the one with the lowest xmin (the one that starts further left).

indices = np.argsort(boxes[0][:,0])

Then you can use these indices to sort the boxes, scores, and classed, as follows:

sorted_scores = scores[0][indices]
sorted_boxes = boxes[0][indices]
sorted_classes = classes[0] indicies

If you wanted to, for example, sort by xmax instead, you'd use np.argsort(boxes[0][:,2]) . You can play around with using 0-3 to sort by xmin, ymin, xmax, and ymax.