finetune_coco

Fine-tune YOLO-World on MS-COCO

Updates

1. [2024-3-27]: Considering that fine-tuning YOLO-World on COCO without mask-refine obtains bad results, e.g., YOLO-World-L obtains 48.6 AP without mask-refine compared to 53.3 AP with mask-refine, we rethink the training process and explore new training schemes for fine-tuning without mask-refine. BTW, the COCO fine-tuning results are updated with higher performance (with mask-refine)!

COCO Results and Checkpoints

NOTE:

1. AP^ZS: AP evaluated in the zero-shot setting (w/o fine-tuning on COCO dataset).
2. mask-refine: refine the box annotations with masks, and add CopyPaste augmentation during training.

model	Schedule	mask-refine	efficient neck	APZS	AP	AP50	AP75	weights	log
YOLO-World-v2-S	AdamW, 2e-4, 80e	✔️	✖️	37.5	46.1	62.0	49.9	HF Checkpoints	log
YOLO-World-v2-M	AdamW, 2e-4, 80e	✔️	✖️	42.8	51.0	67.5	55.2	HF Checkpoints	log
YOLO-World-v2-L	AdamW, 2e-4, 80e	✔️	✖️	45.1	53.9	70.9	58.8	HF Checkpoints	log
YOLO-World-v2-X	AdamW, 2e-4, 80e	✔️	✖️	46.8	54.7	71.6	59.6	HF Checkpoints	log
YOLO-World-v2-L 🔥	SGD, 1e-3, 40e	✖️	✖️	45.1	52.8	69.5	57.8	HF Checkpoints	log

Reparameterized Training

model	Schedule	mask-refine	efficient neck	APZS	AP	AP50	AP75	weights	log
YOLO-World-v2-S	AdamW, 2e-4, 80e	✔️	✖️	37.5	46.3	62.8	50.4	HF Checkpoints	log