Hello,
I have been labelling tables from images that i extracted from PDFs, it is okay but when implementing the jsonl to an image, it is getting more than 320 rows when i only have 10 rows from the image. I understand it is recreating all the scenarios based on the images i labelled. Is there any easier way to handle the labelling? instead of doing it based on the columns and headers, would be better to do it based on the cells?
Hi @giseem5 ,
I'm not quite sure if I understand correctly your workflow. You're using image.manual with images of tables as input. Are the images pre-annotated in any way? Then, you annotate rectangles on images - is that right or is there something else happening on top of that?
Also what do you mean by "when implementing the jsonl to an image"?
Finally:
Is there any easier way to handle the labelling? instead of doing it based on the columns and headers, would be better to do it based on the cells?
I'm really not sure what you mean here. The annotations are spans of pixels, expressed as offsets to a particular point in an image. There's no concept of rows, headers or cells.
You should be getting the same number of examples in and out of recipe (assuming there are no duplicates), and each example should have the same number of spans as there were bounding boxes annotated.
Not sure at which point you are observing the multiplication and to which "rows" you're referring. Since you're annotating tables, I suppose these are the rows in the annotated table. Perhaps it would be easier if you shared the annotated example and the output (from the DB) where you're are seeing the multiplication?