SA-Z: Scene Annotation Dataset with Geometry, Text, and Occlusion Order

A dataset created by FudanCVL, last updated on 2026-05-18. It contains annotations for scene understanding, with each image linked to a global prompt, per-instance geometry and text masks, and pairwise occlusion order. The data is structured as a single JSON file keyed by image_id.

Use Cases

• Training scene understanding models based on the provided global prompts.
• Developing instance segmentation models using the per-instance geometry and text annotations.
• Researching occlusion reasoning and depth ordering from the pairwise occlusion order data.
• Building multimodal models that link visual scenes with textual descriptions.

Strengths

• Provides structured annotations for multiple scene understanding tasks (prompt, mask, order).
• Data is organized in a single, keyed JSON file for programmatic access.
• Last updated on 2026-05-18, indicating recent maintenance.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file size, and license information are unknown, which may limit suitability assessment.

Provenance

Source

FudanCVL via Hugging Face.

Collection Method

Likely collected and annotated for computer vision research.

Time Range

null

Freshness

Last updated 2026-05-18 04:54:09.

Geography

null