Base model: facebook/sam2.1-hiera-small
Base revision: ee5bba1d82bb8749febdf90f45e84b687142ba03
Dataset source: COCO 2017 instance segmentation
Dataset revision: annotations_trainval2017
Split manifest SHA-256: 74dfd6666397ec5398d0757de0b852641a7afc39c6f4ef1a652351653b54d2da
Training samples: 15000
Validation samples: 1000
Test samples: 895
Steps: 350
Batch size: 8
Best validation IoU: 0.782299
Best validation Dice: -1.0
Production output mode: multimask with predicted-IoU candidate selection
Test selected-candidate IoU: 0.800603
Test oracle-candidate IoU: 0.827027
Test selection regret: 0.026425
Test IoU-head MAE: 0.079622
W&B: https://wandb.ai/sshibinthomass-tu-dortmund/household-object-finetuning/runs/5uqfdkuu
Modal app: sam2-household-modal-train-v3
Hub repo: sshibinthomass/sam2-household-accuracy-v3-20260710-175456

Usage

python

from transformers import Sam2Model, Sam2Processor

processor = Sam2Processor.from_pretrained("sshibinthomass/sam2-household-accuracy-v3-20260710-175456")
model = Sam2Model.from_pretrained("sshibinthomass/sam2-household-accuracy-v3-20260710-175456")
outputs = model(**processor(images=image, input_boxes=[[box]], return_tensors="pt"), multimask_output=True)
best_mask_index = outputs.iou_scores[:, 0].argmax(dim=-1)