1
$\begingroup$

There're many object detectors which use convolutional neural networks to detect objects in images or point clouds, like YOLO, SSD, AVOD.

My question is, if i feed the same image to an object detector many times, will the results of each time be the same? Is there any randomness in a CNN object detector that makes it produce different results for the same image?

Thanks in advance.

$\endgroup$
2
  • $\begingroup$ What exactly is the "CNN detector" in your question? $\endgroup$
    – Tim
    Commented Jun 18, 2019 at 18:19
  • $\begingroup$ 'CNN detector' refers to typical object detectors, like YOLO, SSD and those 3D object detectors, which use convolutional neural networks (CNNs) to detect objects in images or point clouds. $\endgroup$
    – Jemma
    Commented Jun 19, 2019 at 1:38

1 Answer 1

1
$\begingroup$

In general, when we want to predict something, we want the predictive algorithm to return consistent results, i.e. same inputs should lead to same outputs. In some cases it may make sense to randomize the predictions (e.g. classification algorithm predicts two classes with equal probabilities, and you need to force choice of a single category), but this is something to be considered on case-by-case basis.

As about neural networks (including convolutional networks), they are defined in terms of algebraic operations (mostly matrix multiplications), activation functions, and some other stuff that is also deterministic. We can use randomization in neural networks, but even if we use things like dropout during training, we turn it off during the prediction phase.

Different story is if we want to simulate the outputs from neural networks using (e.g. autoencoders, GANs, some algorithms in NLP that generate text), then we want them to be randomized.

TL;DR

My question is, if i feed the same image to an object detector many times, will the results of each time be the same?

Yes.

$\endgroup$
1
  • $\begingroup$ Thanks for your answer! $\endgroup$
    – Jemma
    Commented Jun 19, 2019 at 10:52

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.