Reformatting annotations to train recognition model

[zohaib-khan5040]

Hello!

I have a dataset with annotations similar to the YOLO format - for each image I have a list of bounding boxes, each having coordinates for x,y, and the width and height. I wish to make annotations similar to what is provided in the references/ folder for finetuning detection and recognition models.

I can get the detection models to work, that's not an issue. I do want to ask how I can manipulate these annotations to get a structure for the recognition dataset. I was initially thinking of getting a reading order for the boxes but I doubt that's a good way to go about it.

Thank you! 🙏

[zohaib-khan5040]

Found this function in the repo:

doctr/doctr/models/builder.py

Line 43 in 95857cf

def _sort_boxes(boxes: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:

Still curious if this is the best approach for something like this.

[felixdittrich92]

Hi @zohaib-khan5040 👋,

Not sure what you want to reach 😅 but the order doesn't matter if you want to train a recognition model (it's enough to provide the Image crops and coresponding labels)
So could you explain your problem a bit more please ? :)