Datasets:
Tasks:
Visual Question Answering
Formats:
parquet
Languages:
English
Size:
10K - 100K
ArXiv:
Tags:
remote-sensing
satellite-imagery
geospatial
earth-observation
spatial-reasoning
quantitative-reasoning
DOI:
License:
Simplify dataset card
Browse files
README.md
CHANGED
|
@@ -29,15 +29,15 @@ configs:
|
|
| 29 |
<img src="https://huggingface.co/datasets/PeterAM4/SQuID/resolve/main/assets/logo.png" width="90">
|
| 30 |
</p>
|
| 31 |
|
| 32 |
-
<h1 align="center">SQuID
|
| 33 |
|
| 34 |
<p align="center">
|
|
|
|
| 35 |
<a href="https://arxiv.org/abs/2601.13401"><img src="https://img.shields.io/badge/arXiv-2601.13401-b31b1b.svg"></a>
|
| 36 |
<a href="https://github.com/PeterAMassih/qvlm-squid"><img src="https://img.shields.io/badge/GitHub-code-181717.svg?logo=github"></a>
|
| 37 |
-
<a href="https://doi.org/10.57967/hf/7565"><img src="https://img.shields.io/badge/DOI-10.57967%2Fhf%2F7565-blue.svg"></a>
|
| 38 |
</p>
|
| 39 |
|
| 40 |
-
SQuID
|
| 41 |
|
| 42 |

|
| 43 |
|
|
@@ -46,42 +46,36 @@ SQuID tests whether vision-language models can **measure** satellite images: cou
|
|
| 46 |
```python
|
| 47 |
from datasets import load_dataset
|
| 48 |
|
| 49 |
-
bench = load_dataset("PeterAM4/SQuID", "benchmark", split="test")
|
| 50 |
-
train = load_dataset("PeterAM4/SQuID", "train", split="train")
|
| 51 |
```
|
| 52 |
|
| 53 |
-
A numeric answer is correct
|
| 54 |
|
| 55 |
-
Training
|
| 56 |
|
| 57 |
```python
|
| 58 |
from huggingface_hub import snapshot_download
|
| 59 |
-
snapshot_download("PeterAM4/SQuID", repo_type="dataset", local_dir="data"
|
| 60 |
-
allow_patterns=["benchmark/*", "train/*"])
|
| 61 |
```
|
| 62 |
|
| 63 |
## Data
|
| 64 |
|
| 65 |
-
| Split | Questions | Images |
|
| 66 |
-
|---|---:|---:|
|
| 67 |
-
| `benchmark` | 2,000 | 639 |
|
| 68 |
-
| `train` | 75,687 | 1,443 |
|
| 69 |
|
| 70 |
-
- **
|
| 71 |
-
- **
|
| 72 |
-
- **Tolerances** come from
|
|
|
|
| 73 |
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
**Masks.** DeepGlobe `*_label.png` and EarthVQA `*_label2.png` are RGB class maps, and Solar `*_label.bmp` is binary. EarthVQA's human labels leave about a quarter of each tile as background. `_label2` fills that background with a land-cover model's predictions (`_pred2`), and the original label is kept as `_label`.
|
| 77 |
-
|
| 78 |
-
## Versions
|
| 79 |
-
|
| 80 |
-
**v2.0** (current) adds the training split and ground-truth masks. It also computes areas as exact pixel counts with holes excluded, defines vegetation in the question text, and removes questions whose answer was a forced zero. **v1.0** is the January 2026 benchmark (`revision="v1.0"`, [DOI 10.57967/hf/7565](https://doi.org/10.57967/hf/7565)).
|
| 81 |
|
| 82 |
## License and citation
|
| 83 |
|
| 84 |
-
|
| 85 |
|
| 86 |
```bibtex
|
| 87 |
@misc{massih2026reasoningpixellevelprecisionqvlm,
|
|
|
|
| 29 |
<img src="https://huggingface.co/datasets/PeterAM4/SQuID/resolve/main/assets/logo.png" width="90">
|
| 30 |
</p>
|
| 31 |
|
| 32 |
+
<h1 align="center">SQuID</h1>
|
| 33 |
|
| 34 |
<p align="center">
|
| 35 |
+
Quantitative spatial reasoning on satellite imagery<br><br>
|
| 36 |
<a href="https://arxiv.org/abs/2601.13401"><img src="https://img.shields.io/badge/arXiv-2601.13401-b31b1b.svg"></a>
|
| 37 |
<a href="https://github.com/PeterAMassih/qvlm-squid"><img src="https://img.shields.io/badge/GitHub-code-181717.svg?logo=github"></a>
|
|
|
|
| 38 |
</p>
|
| 39 |
|
| 40 |
+
SQuID asks vision-language models to measure satellite images: count objects, compute areas and reason about distances. Every answer is computed from pixel-level masks.
|
| 41 |
|
| 42 |

|
| 43 |
|
|
|
|
| 46 |
```python
|
| 47 |
from datasets import load_dataset
|
| 48 |
|
| 49 |
+
bench = load_dataset("PeterAM4/SQuID", "benchmark", split="test")
|
| 50 |
+
train = load_dataset("PeterAM4/SQuID", "train", split="train")
|
| 51 |
```
|
| 52 |
|
| 53 |
+
A numeric answer is correct inside `[acceptable_range_lower, acceptable_range_upper]`. Binary and `connected`/`fragmented` answers must match exactly.
|
| 54 |
|
| 55 |
+
Training rows point to image files. Download them, with all masks and raw JSON files:
|
| 56 |
|
| 57 |
```python
|
| 58 |
from huggingface_hub import snapshot_download
|
| 59 |
+
snapshot_download("PeterAM4/SQuID", repo_type="dataset", local_dir="data")
|
|
|
|
| 60 |
```
|
| 61 |
|
| 62 |
## Data
|
| 63 |
|
| 64 |
+
| Split | Questions | Images |
|
| 65 |
+
|---|---:|---:|
|
| 66 |
+
| `benchmark` | 2,000 | 639 |
|
| 67 |
+
| `train` | 75,687 | 1,443 |
|
| 68 |
|
| 69 |
+
- **Sources:** DeepGlobe (0.5 m), EarthVQA (0.3 m), solar panels (0.3 m), and 50 human-answered NAIP questions (1.0 m) in the benchmark. The two splits share no images.
|
| 70 |
+
- **Tiers:** basic (713), spatial (610) and multi-step (677) questions, of 25 types.
|
| 71 |
+
- **Tolerances** come from 500 human annotations: ±1.7 points for percentages and ±19% for counts.
|
| 72 |
+
- **Masks:** EarthVQA answers use `_label2`, the human label with its background filled by a land-cover model (`_pred2`).
|
| 73 |
|
| 74 |
+
v1.0, the January 2026 benchmark, is still available with `revision="v1.0"` ([DOI](https://doi.org/10.57967/hf/7565)).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 75 |
|
| 76 |
## License and citation
|
| 77 |
|
| 78 |
+
CC BY-NC 4.0. Images keep the terms of [DeepGlobe](https://arxiv.org/abs/1805.06561), [EarthVQA](https://github.com/Junjue-Wang/EarthVQA), [PV segmentation](https://doi.org/10.5281/zenodo.5171712) and [NAIP](https://naip-usdaonline.hub.arcgis.com/).
|
| 79 |
|
| 80 |
```bibtex
|
| 81 |
@misc{massih2026reasoningpixellevelprecisionqvlm,
|