I read in many paper that mentions coarse-to-fine as a technique in deep learning, but I could never figure what exactly they mean. Is it related to multiscale inference, where they use coarse and fine input images?

  • 1
    $\begingroup$ This question is probably incomplete , was part of it lost in copy/paste ? $\endgroup$ Mar 25, 2019 at 6:56
  • $\begingroup$ Opps I don't know what happened. Completing it now. $\endgroup$
    – Mong H. Ng
    Mar 26, 2019 at 7:14

1 Answer 1


"Coarse to Fine" usually refers to the hyperparameter optimization of a neural network during which you would like to try out different combinations of the hyperparameters and evaluate the performance of the network.

However, due to the large number of parameters AND the big range of their values, it is almost impossible to check all the available combinations. For that reason, you usually discretize the available value range of each parameter into a "coarse" grid of values (i.e. val = 5,6,7,8,9) to estimate the effect of increasing or decreasing the value of that parameter. After selecting the value that seems most promising/meaningful (i.e. val = 6), you perform a "finer" search around it (i.e. val = 5.8, 5.9, 6.0, 6.1, 6.2) to optimize even further.

  • $\begingroup$ That's a good answer for the amount of information given. Since he was talking about images there are some methods that train multiple regressors for coarse-to-fine detection of objects. I find this a lot in facial landmark detection. So coarse-to-fine might mean: - Oposite to exhaustive grid search or - Oposite to one shot-detector / yolo $\endgroup$ Mar 26, 2019 at 11:31

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.