PeterAM4 commited on
Commit
867cc22
·
verified ·
1 Parent(s): eda516c

Simplify dataset card

Browse files
Files changed (1) hide show
  1. README.md +18 -24
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: Satellite Quantitative Intelligence Dataset</h1>
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 tests whether vision-language models can **measure** satellite images: count objects, compute areas in hectares, and reason about distances. Every answer is computed from pixel-level masks, and numeric answers are scored against tolerance ranges calibrated on human annotators.
41
 
42
  ![Examples](https://huggingface.co/datasets/PeterAM4/SQuID/resolve/main/assets/examples.jpg)
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") # 2,000 questions, images + masks
50
- train = load_dataset("PeterAM4/SQuID", "train", split="train") # 75,687 questions, file paths
51
  ```
52
 
53
- A numeric answer is correct if it falls in `[acceptable_range_lower, acceptable_range_upper]`. Binary (`0`/`1`) and `connected`/`fragmented` answers must match exactly.
54
 
55
- Training images and masks are stored as files. Download everything, including the raw JSON files used by the [code](https://github.com/PeterAMassih/qvlm-squid), with:
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 | Sources (GSD) |
66
- |---|---:|---:|---|
67
- | `benchmark` | 2,000 | 639 | DeepGlobe (0.5 m) 549 · EarthVQA (0.3 m) 1,304 · Solar (0.3 m) 97 · NAIP (1.0 m) 50 |
68
- | `train` | 75,687 | 1,443 | DeepGlobe 35,542 · EarthVQA 36,190 · Solar 3,955 |
69
 
70
- - **Tiers:** 1 basic (coverage, counts, sizes), 2 spatial (proximity, connectivity, power), 3 multi-step (filter by size, then clip by distance). The benchmark has 713 / 610 / 677 questions per tier and 25 question types.
71
- - **Splits** are image-disjoint and come from the official train / evaluation partitions of each source. The 50 NAIP questions are answered by humans (median of 10 annotators).
72
- - **Tolerances** come from the disagreement of 500 human annotations: ±1.735 points for percentages, ±2.25 for proximity percentages, ±max(1, 0.19·C) for counts, and ±1.735 % (±2.25 % near features) of the value for areas.
 
73
 
74
- **Fields.** `benchmark`: `id`, `image`, `mask` (none for NAIP), `question`, `answer` (string), `type`, `tier`, `gsd`, `acceptable_range_lower`, `acceptable_range_upper`, `source`. `train`: the same text fields, with `image` and `mask` as paths into `train/images_and_masks/`.
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
- Questions, answers and annotations are released under CC BY-NC 4.0. Images and masks are redistributed for non-commercial research from [DeepGlobe](https://arxiv.org/abs/1805.06561), [EarthVQA](https://github.com/Junjue-Wang/EarthVQA), the [PV segmentation dataset](https://doi.org/10.5281/zenodo.5171712) and [NAIP](https://naip-usdaonline.hub.arcgis.com/), and remain subject to their original terms.
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
  ![Examples](https://huggingface.co/datasets/PeterAM4/SQuID/resolve/main/assets/examples.jpg)
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,