SAM 2.1 Hiera Tiny - Promptable Instance Masks

Exact checkpoint

The selected file is sam2.1_hiera_tiny.pt from Meta's improved SAM 2.1 checkpoint release. Pair it with configs/sam2.1/sam2.1_hiera_t.yaml. This is a deliberately selected small SAM 2.1 variant; no claim is made that it is the newest Segment Anything generation.

How it works

An image encoder creates reusable features. A prompt encoder represents clicks or boxes; the decoder combines them to produce masks. Video use additionally carries memory between frames. Automatic mask generation can propose many regions, but it is not a semantic classifier.

Typical input and output

Input is RGB imagery plus optional point or box prompts through SAM2ImagePredictor. The selected configuration specifies image_size 1024. Outputs include candidate masks and mask-quality scores; use the predictor to map masks back to the original image size.

Strengths

Interactive prompts give an analyst direct control over which object to outline. This is useful for annotation assistance and iterative mask correction.

GIS and remote-sensing use

Use as an assisted digitizing tool: pick an object, inspect the mask, correct it and only then attach a semantic label. Preserve the chip transform when turning approved masks into GIS polygons. Keep a review step for survey, cadastral or safety-sensitive boundaries.

Limitations

A returned mask has no guaranteed building, tree or water label. Adjacent crowns, shadows and very small objects may need additional prompts or manual correction. Video tracking does not establish reliability for seasonally separated satellite dates.

Training and fine-tuning

For adaptation, use the official training workflow with image or video masks and appropriate prompts. Keep object identities consistent across video frames. Compare adaptation against the unmodified prompted baseline on a geographically separate area, measuring mask quality and the amount of human correction required.

Framework and hardware

Official PyTorch SAM 2 implementation. The repository documents Python 3.10+, PyTorch 2.5.1+ and TorchVision 0.20.1+, and recommends Linux/WSL for Windows. Its CUDA-oriented setup and optional postprocessing need compatible hardware; no universal CPU or VRAM promise is made.

Files, license and download

Meta licenses checkpoints under Apache-2.0. The official file is 156,008,466 bytes, above GISSchools' existing 100 MiB resource ceiling, so the genuine provider download is linked.

Before using a result

1. Keep imagery rights, acquisition dates and preprocessing with each experiment.
2. Separate locations before tiling to avoid train/test leakage.
3. Inspect failures and uncertain outputs alongside the original imagery.
4. Preserve coordinate transforms and verify units before exporting a GIS layer.

Validation status: Source provenance and website delivery are checked. GISSchools has not run model inference, training or an accuracy benchmark for this record. Feature images and gallery diagrams are explanatory illustrations, not predictions.

Official references

• Original project
• Usage documentation
• Research paper
• Selected official checkpoint
• License and usage conditions