SAMRS: Large-Scale Remote Sensing Segmentation via Segment Anything Model

SAMRS provides remote sensing segmentation data generated using the Segment Anything Model (SAM), published by ViTAE-Transformer for NeurIPS 2023. The dataset scales up available masks for remote sensing imagery to facilitate large-scale pre-training and transfer learning in the geospatial domain.

Use Cases

• Pre-training vision transformers using SAM-generated segmentation masks
• Transfer learning for remote sensing semantic segmentation tasks
• Evaluating Segment Anything Model zero-shot performance on satellite imagery

Strengths

• Peer-reviewed methodology from NeurIPS 2023
• Utilizes Segment Anything Model (SAM) for high-volume mask generation
• Specifically optimized for remote sensing pre-training

Limitations

• Masks are model-generated (SAM) rather than expert-annotated, introducing potential label noise
• Lack of human verification for every generated segment

Provenance

Source

ViTAE-Transformer (NeurIPS 2023)

Collection Method

Automated annotation using the Segment Anything Model (SAM) applied to existing remote sensing image archives.

Freshness

Last updated August 2024; reflects research published in late 2023.