Finetuning YOLOV5-OBB

[rohanchn]

Hi @PonteIneptique,

I did some training with the usual v5x6 model, and initial tests suggests very good results. Here's an example of assesment during validation.

However, I also think my data may benefit from rotating bounding boxes (as deskew on two facing pages in one image could be a problem, and also the way GraphicZone appear in my data), so thought I'd also test YOLOv5-OBB that you mention in the Yaltai paper.

I have a couple of questions:

1. How do I get the YOLOv5-OBB model?

2. Do you think a baseline model trained in ketos segtrain with --suppress-regions could be more efficient than --no-suppress-regions for a Yaltai > Kraken pipeline?

[PonteIneptique]

Hi ^^

So, first of all, to address one elephant in the room, as lines are anchored based on their center, deskewing is mostly useless if what you are looking into is have you line correctly ordered in different columns.

1. YOLO OBB is here: https://github.com/hukaixuan19970627/yolov5_obb/ but I would have to adapt the Alto to YOLO to produce oriented box.
2. Yes, it simply produces less to postprocess :)

[rohanchn]

Right, I thought so too.

Thanks! I got the same link but couldn't get myself to trust what I was looking, so posted here.

[PonteIneptique]

FYI, I am looking to implement yolo8 models but I need to check before if they work well :)

[rohanchn]

Isn't that the latest release? Sounds something I will look forward to try!

[PonteIneptique]

YALTAi does not use this yet. But it might. I need to benchmark and test :)

[rohanchn]

pip install YALTAi doesn't reflect the latest changes vis-a-vis manifest. You may want to tag a new release.

[PonteIneptique]

@rohanchn You might be interested in knowing I moved YALTAi to YOLOv8 which allows for polygons. I have not tested polygon yet (and YALTAi does not allows for Polygons yet), but it might be a solution for this :)

[rohanchn]

Thank you! I should be able to test soon.