2
$\begingroup$

I need to build a hand detector that recognizes the chord played by a hand on a guitar.

I read this article Static Hand Gesture Recognition using Convolutional Neural Network with Data Augmentation that looks like what I need (hand gesture recognition).

I think my task is (from my point of view) a little more difficult than that in the paper, because I think it is more difficult to distinguish between two chords than between a punch and a palm.

What I don't understand clearly is how to choose the best parameters for this more complex task: is it better to have more/less convolutional layers? A higher or lower number of poolings? Max or avg pooling?

The input will be more or less like this one:

Input

There will be a first net (MobileNetV2 trained on EgoHands) that will find the bounding box, crops the image and then passes the saturated blending between the original one and Frei&Chen edges to the second net (unfortunately I don't have a processed picture yet, I will post an example as soon as I get it)

$\endgroup$
1
  • 1
    $\begingroup$ there is no way to compute analytically what the best parameters are. It is also not trivial to conclude that the task is more complex simply cause for "us" is more difficult to distinguish between 2 chords than palm and punch. It all depends on the data you have. Any concrete picture to share with? Or idea on how you would collect the data in terms of camera, position, lighting, angle? $\endgroup$ Oct 8 '21 at 9:12
0
+50
$\begingroup$

The ideal hyperparameters is usually dependent on your dataset and will differ on a case by case basis. Go for trial and error to determine the hyperparameters that works best for you.

Few research papers similar to your use case is listed below.

$\endgroup$
0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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