Sahibnoor1/kvasir-siglip2-segmentation-checkpoints

Kvasir-SEG SigLIP2 Segmentation Checkpoints

This repository contains checkpoints and evaluation artifacts for polyp segmentation on Kvasir-SEG.

Models

• ResUNet reproduction
• SigLIP2 frozen encoder
• SigLIP2 partial fine-tuning
• SigLIP2 full fine-tuning

Each method was trained with seeds 42, 43 and 44.

Main results

Method	Mean Dice	Seed SD	Mean IoU	Precision	Recall
ResUNet	0.8182	0.0072	0.7326	0.8739	0.8354
SigLIP2 Frozen	0.8745	0.0010	0.8024	0.9015	0.8903
SigLIP2 Partial	0.8813	0.0039	0.8116	0.8943	0.9072
SigLIP2 Full	0.9013	0.0026	0.8447	0.9170	0.9211

Full SigLIP2 versus ResUNet:

• Mean Dice improvement: +0.0832
• Paired bootstrap 95% CI: [0.0546, 0.1147]
• Paired permutation test: p < 0.001
• Official test split size: 120 images

Checkpoint structure

checkpoints/
├── resunet_paper/
├── siglip2_frozen/
├── siglip2_partial/
└── siglip2_full/