So, in the above example, it seems you got 100 detections.

Each row in the above contains the normalised coordinates in the following form[x_min, y_min, x_max, y_max]. To get the actual coordinates, you need to multiplyx_min andx_max of each detection by the width of the image, and, similarly,y_min andy_max by the height of the image.

Also, make sure to convert the results to integer values.

well, I have only one object that has been detected but may be it looks like this because of this line>>

The model expects a batch of images, so add an axis withtf.newaxis.

input_tensor = input_tensor[tf.newaxis, ...]

Also, acoording to the comments in the above code we are only intrested in the first num_detection so I took the first array >>

[0.36031908 0.5819093 0.6042158 0.99556196]

and then multiplied it by the width and height then I used the result to cut the detected object by this method you provided:https://www.geeksforgeeks.org/how-to-crop-an-image-using-the-numpy-module/

but the cut is not correct, do you have another cropping method using x_min y_min x_max y_max?

Hmmm, that's interesting.

When you sliced the array, did you do it as follows?

cropped = original[y_min:y_max, x_min:x_max]

Note:x_min, y_min, etc here refer to the unnormalised coordinates.

Ahhh, I just noticed I may have made a mistake. The coordinates returned by the detector are of the form[y_min, x_min, y_max, x_max], and not[x_min, y_min, x_max, y_max]!

Thank you so much. Finally I got what I want.