fine-tuning panoptic segmentation

WongKinYiu / yolov9

Implementation of paper - YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information

GNU General Public License v3.0

8.65k stars 1.32k forks source link

fine-tuning panoptic segmentation #243

Open haujulian opened 4 months ago

haujulian commented 4 months ago

I would like to fine-tune gelan-c-pan.pt on a custom dataset. However I do not quite understand how to adjust the panoptic head. If I understand gelan-c-pan.yaml correctly ([[15, 18, 21], 1, Panoptic, [nc, 93, 32, 256]], # Panoptic(P3, P4, P5)), nc refers to the number of thing classes and 93 refers to the number of stuff classes? Do I need to list thing and stuff classes separately in the data.yaml file for the training to work? Thanks for any help!

WongKinYiu commented 4 months ago

coco.yaml updated.

haujulian commented 4 months ago

When generating the .txt files for stuff and things, are the class-ids allowed to overlap between stuff and things? Do the class-ids of stuff in the .txt files need to match the indices of stuff_names? I also think panoptic/predict.py is not working correctly with gelan-c-pan.pt.

RohitKeshari commented 3 months ago

How to switch off stuff? My custom data only has class polygon