SA-Co/Silver: A Promptable Concept Segmentation Benchmark Across 10 Domains

SA-Co/Silver is a benchmark for promptable concept segmentation in images, created by Facebook and last updated in November 2025. It contains images paired with text labels, each exhaustively annotated with masks for all object instances matching the label. The benchmark comprises 10 subsets covering diverse domains including food, art, robotics, and driving.

Use Cases

• Benchmarking promptable segmentation models based on the exhaustive mask annotations for text labels.
• Training vision-language models for concept grounding based on the image-text-mask triplets.
• Evaluating model generalization across domains based on the 10 diverse subsets mentioned in the description.
• Researching instance-level segmentation and recognition based on the annotated object instances.

Strengths

• Exhaustive annotation of masks for all object instances matching a given text label.
• Diverse coverage across 10 domains including food, art, robotics, and driving.
• Created by Facebook, a major AI research organization.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Specific license terms are not provided in the input.

Provenance

Source

Facebook

Collection Method

Likely aggregated and annotated from multiple public datasets including BDD100k, DROID, and Ego4D.

Time Range

null

Freshness

Last updated 2025-11 17:53:40.

Geography

null