Question about tf.image.non_max_suppression

In tf.image.non_max_suppression, where the documentation says that:

this algorithm is invariant to orthogonal transformations and translations of the coordinate system

Does this means that the result shouldn’t change if the input is given as [x1,y1,x2, y2] instead of the required [y1, x1, y2, x2]?

Yes, you are correct, the result will remain unchanged if the axes get swapped.