📌 Introduction
Scene graph generation (SGG) aims to understand the visual objects and their semantic relationships from aerial images. While many SGG datasets exist for eye-level views, overhead views are under-explored. To address this, we introduce UASG: a new urban aerial scene graph dataset with 25,594 objects, 16,970 relationships, and 27,175 attributes.
We further propose a novel Locality-Preserving Graph Convolutional Network (LPG) that effectively embeds object features with scene-level spatial context, while pruning redundant relationship pairs using an adaptive bounding box scaling strategy.
📊 Dataset Statistics
| Type | Objects | Relationships | Attributes |
|---|---|---|---|
| Total | 25,594 | 16,970 | 27,175 |
🖼️ Example Images
