Getting started: YOLOX inference and image preprocessing about super-gradients HOT 4 CLOSED

@shayamram If you want to use the same preprocessing that was used for validation- then yes padded rescale to 640X640 with BGR.
The dog was a square image anyway and the model handled it.
You can see the exact preprocessing that is used for validation for any recipe if you check their dataset_params, for instance these are the dataset_params that were used for training YoloX, and
notice the transforms part inside valid_dataset_params.

from super-gradients.

Join the discussion on DagsHub!

from super-gradients.

Hey @shayamram (:

I assume you are pointing to the "Predict" cell, notice the line:

image_tensor = torch.tensor(np.array(image)[:, :, ::-1].copy()).permute(2, 0, 1).unsqueeze(dim=0).float().cuda()

The [:, :, ::-1] part reverses the order of the 3rd dimension, and therefore converts it to BGR format.
I will add a comment to clarify that this is also being done there.

If you have any other questions, please let me know.

from super-gradients.

@shaydeci Thanks for you reply - so just to clarify: the model predict expects BGR and padded image? (example: the image was padded, I assume to maintain aspect ratio).

This seems different from the example here (the one with the dog). This example is RGB and only simple resize (no padding), so I just want to be extra-certain that the correct preprocessing for prediction is (resize+padding) in BGR, and not simple-resize (no padding) RGB as in the second example.

Thanks again

from super-gradients.