RS-Spatial

Training data for Uni-RS: A Spatially Faithful Unified Understanding and Generation Model for Remote Sensing (IJCAI 2026). [Paper] [Code]

Archive	Content	Rows
layout_plans.zip	<PLAN> layout plans for Spatial-Layout Planning	22,849
spatial_relations.zip	Spatial relation labels (K = 79) for Spatial-Aware Query Supervision	22,849
rotation_pairs.zip	Rotated images with matching captions for Image-Caption Spatial Layout Variation	36,702

Files

Unzip the three archives into one directory:

rs_spatial/
├── layout_plans/
│   ├── rsicap.jsonl
│   └── vrsbench.jsonl
├── spatial_relations/
│   ├── rsicap.jsonl
│   ├── vrsbench.jsonl
│   └── spatial_vocabulary.json
└── rotation_pairs/            Hugging Face dataset (save_to_disk format)
    ├── data-00000-of-00026.arrow
    ├── ...
    ├── dataset_info.json
    └── state.json

layout_plans/{rsicap,vrsbench}.jsonl

One JSON object per line. Rows are aligned by index with the RSICap and VRSBench caption datasets.

Field	Type	Description
idx	int	Row index in the caption dataset
caption	str	Caption
plan	str	Layout plan enclosed in <PLAN> ... </PLAN>

A plan contains SCENE (global scene), OBJ (object instances with a coarse position on a 3x3 grid: TopLeft, TopCenter, ..., BottomRight, plus optional count and attribute) and REL (relations between objects) lines:

<PLAN>
SCENE desc="rural area" source="GoogleEarth"

OBJ o1 type="expressway toll station" pos=MiddleCenter
OBJ v1 type="small vehicle" pos=MiddleLeft
OBJ v2 type="small vehicle" pos=MiddleCenter desc="right-most relative to the toll station"
OBJ v3 type="small vehicle" count=multiple desc="alongside the toll station"

REL Beside(v1, o1)
REL Beside(v2, o1)
REL Beside(v3, o1)
</PLAN>

spatial_relations/{rsicap,vrsbench}.jsonl

One JSON object per line. Rows are aligned by index with the RSICap and VRSBench caption datasets.

Field	Type	Description
idx	int	Row index in the caption dataset
caption	str	Caption
spatial_relations	list[str]	Spatial relation predicates present in the caption
spatial_labels	list[int]	79-dimensional multi-hot vector over the vocabulary

spatial_vocabulary.json lists the 79 predicates in groups by arity (unary / binary) and dimension (directional, topological, distance, orientation, distribution). Label index k is the k-th predicate when the groups are read in order.

rotation_pairs/

A Hugging Face datasets dataset with 36,702 rows.

Field	Type	Description
id	str	{source_dataset}:{source_idx}:{rotation_angle}
image	Image	512 x 512 PNG, rotated clockwise by rotation_angle
caption	str	Caption matching the rotated image
rotation_angle	int	90, 180 or 270
original_caption	str	Caption of the unrotated image
source_dataset	str	rsicap or vrsbench
source_idx	int	Row index in the caption dataset
source_key	str	Identifier of the source image

from datasets import load_from_disk

ds = load_from_disk("rs_spatial/rotation_pairs")
print(ds[0]["caption"], ds[0]["rotation_angle"])
ds[0]["image"].show()

Usage with Uni-RS

Place rs_spatial/ under UNIRS_DATA_ROOT next to the RSICap and VRSBench datasets and run Stage 2 training as described in the code repository.

License

CC BY-NC 4.0. Images and captions from RSICap and VRSBench are subject to their own licenses.

Citation

@inproceedings{ijcai2026p231,
  title     = {Uni-RS: A Spatially Faithful Unified Understanding and Generation Model for Remote Sensing},
  author    = {Zhang, Weiyu and Hu, Yuan and Li, Yong and Liu, Yu},
  booktitle = {Proceedings of the Thirty-Fifth International Joint Conference on Artificial Intelligence, {IJCAI-26}},
  pages     = {2073--2081},
  year      = {2026},
  doi       = {10.24963/ijcai.2026/231},
}