SAM 2.1 base-plus, HyperKD-RS fine-tuned decoder

Mask decoder + prompt encoder of SAM 2.1 hiera-base-plus fine-tuned on cp500/hyperkd-rs: point prompt -> track-consensus mask over 38 Landsat spot shards (1985-2025, 30 m). Encoder frozen.

Held-out (temporal holdout, last 2 scenes/spot, argmax-IoU head): zero-shot 0.365 -> fine-tuned 0.565 IoU vs consensus targets. 4,000 steps, BCE+Dice+IoU-head MSE, best-head supervision, lr 1e-5 AdamW.

Load: build SAM 2.1 b+ from the official repo/checkpoint, then load this state dict into sam.sam_mask_decoder / sam.sam_prompt_encoder. Training script: harness/finetune_sam21.py in the dataset repo.

Downloads last month
49
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for cp500/sam21-hyperkd-rs

Quantized
(5)
this model

Space using cp500/sam21-hyperkd-rs 1