2
$\begingroup$

From tensorflow's object recognition (R-CNN)

I'm re-training the existing model with new categories: the types of clothes (jeans, pants, blouse, and so on). Since we don't need colors to determine the type of clothes that user is wearing, I want to re-train it with gray-scale images. Is it possible to use gray-scale images to train existing model (which are trained with color images)?

I'm concerned because they trained their model with color images.

Does the model just consider the grayscale image as color image? And does it still work? :)

p.s I'm generating XML and csv files to put data for training and testing.

$\endgroup$
1
$\begingroup$

It depends on the model. You'll have to dig into the model definition and see.

It's not uncommon for models to grayscale images in pre-processing. In that case, you will be fine.

However, many recent deep convolutional models operate on RGB images as 3D tensors (4D with the batch dimension). In this case, you should consider modifying the model so that it operates on a single color channel before training it. Multi-channel convolution is discussed in the convent's chapter of Goodfellow's Deep Learning.

$\endgroup$

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.