STAR: Large‑Scale Satellite Image Scene‑Graph Generation Dataset and Benchmark

Overview

STAR is the first large‑scale dataset for scene‑graph generation (SGG) on very‑high‑resolution (VHR) satellite imagery. It comprises over 210,000 objects and 400,000 relational triples, covering 1,273 diverse global scenes.

Dataset Characteristics

• Scale: >210,000 objects and 400,000 triples.
• Image Sizes: From 512 × 768 px up to 27,860 × 31,096 px.
• Complexity: Objects vary widely in scale and aspect ratio, with rich inter‑object relationships.

Construction

• Goal: Promote geographic scene understanding from perception to cognition.
• Approach: Introduces a Context‑Aware Cascaded Cognition (CAC) framework for object detection (OBD), pair pruning, and relation prediction.

Dataset & Toolkit

• Links: STAR Dataset and Toolkit

Usage

• Installation & Usage: Detailed installation, pretrained models, training, and evaluation instructions are provided in the MMRotate 0.3.4 repository.

Published Models

Oriented Object Detection

Below are several released oriented object detection models and their performance metrics:

(The remainder of the table is omitted for brevity; all entries retain the original markdown structure.)