Datasets:
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset contains images of human subjects and is released for non-commercial research use only. By requesting access you agree to: (1) use the data solely for research purposes; (2) not redistribute the images or annotations; (3) not attempt to identify any person appearing in the images; (4) delete the data upon request of the dataset maintainers.
Log in or Sign Up to review the conditions and access this dataset content.
Bench-SPACE
Bench-SPACE is the dataset of the SPACE benchmark from "Human-Centric Image Captioning with Subject-Centered Spatial Understanding". It evaluates how well vision-language models describe human subjects in images, focusing on Spatial & Orientation, Action & Pose, Appearance, and Multi-person Interaction. Model captions are scored by an LLM judge against per-image QA pairs derived from human-verified descriptions.
- 2,981 images (~4.3 GB)
- 19,892 QA pairs across 16 categories
Evaluation code: github.com/JHang2020/SPACE-Eval
Files
SPACE.jsonl # one QA pair per line
images/ # 2,981 images, referenced by relative path
Annotation schema (SPACE.jsonl)
| field | description |
|---|---|
qa_id |
unique QA identifier |
raw_id |
image file name |
image_path |
image path relative to this file, e.g. images/xxx.JPEG |
category_id |
question category (see below) |
question |
question about the human subject(s) |
ground_truth |
reference answer derived from a human-verified description |
translated_caption_en |
English translation of the source human caption |
origin_caption_cn |
original human-written Chinese caption |
Category distribution
| category_id | #QA | category_id | #QA |
|---|---|---|---|
| 1.1 | 3,304 | 3.1 | 630 |
| 1.2 | 759 | 3.2 | 630 |
| 1.3 | 927 | 3.3 | 630 |
| 1.4 | 1,056 | 3.4 | 566 |
| 1.5 | 1,177 | 3.5 | 909 |
| 2.1 | 2,145 | 3.6 | 979 |
| 2.2 | 2,145 | 3.7 | 1,120 |
| 2.3 | 2,145 | 3.8 | 770 |
Usage
huggingface-cli download <HF_DATASET_REPO_ID> --repo-type dataset --local-dir anno_data
Then follow the evaluation pipeline in the code repository: github.com/JHang2020/SPACE-Eval.
Ethics & privacy
All images contain human subjects. Access is gated and restricted to non-commercial research. Do not redistribute the images or attempt to identify individuals. For takedown requests, contact the dataset maintainers.
Citation
TBD
- Downloads last month
- 12