Doctr: Predicted Boundary Boxes vs Target Boundary Boxes

[juanpajedrez]

Hi everyone!

Recently, I started to create a project to perform Optical character recognition, and I started using docTR in order to experiment and see how text detection and text recognition can be made.

So far, I have being looking into the FUNSD dataset (since that's the kind of data I am working with).

These are the main codes I used in Pytorch in order to obtain the resullts.

However... this is where it gets funky.

When I took a look into a specific output, (I understand the N x 5 is equivalent to N x 4 dimensions + a confidence score), I see my height and width calculations to be ok.... but the predicted pixel.... is just too off. Plotting green target vs red predicted showcases that way more.

Am I doing something wrong with my PyTorch?