🕐 TickTockVQA

Analog clock reading dataset from It's Time to Get It Right: Improving Analog Clock Reading and Clock-Hand Spatial Reasoning in Vision-Language Models (CVPR 2026 Findings). Contains 12,483 images with time labels for training and evaluating VQA models on analog clock understanding.

Dataset Notice: This dataset is collected from publicly available data corpora. The copyright and redistribution conditions of the dataset do not belong to the authors of this project. Please refer to the respective source data corpora and the license information stored in the dataset columns for details on usage, attribution, and redistribution terms.

---

📖 Overview

TickTockVQA is a benchmark dataset for reading analog clocks from images. Each sample includes an image and the ground-truth time displayed on the most prominent clock face.

Split	Samples
Train	7,236
Test	5,247

Sources: OpenImages, COCO, ClockMovies, VisualGenome, CC12M, ImageNet, SBU

---

🚀 Quickstart

from datasets import load_dataset
import matplotlib.pyplot as plt

dataset = load_dataset("jaeha-choi/TickTockVQA")
print(dataset)

sample = dataset["train"][0]
print(sample["time_string"])
print(sample["image"])

plt.imshow(sample["image"])
plt.axis("off")
plt.show()

The image column decodes to a PIL.Image.Image object by default. The original relative reference is preserved in image_path.

Schema

All annotation metadata from the old annotations.json file is preserved as dataset columns, with two additional fields:

• split: explicit split name derived from image_path
• image: embedded image data stored as a Hugging Face Image feature

Column	Type
ampm	string
attribution_required	bool
clock_type	string
commercial_use	bool
design	list[string]
environment	string
hour	int64
image_hash	string
image_name	string
image_path	string
license	string
license_notes	string
license_type	string
license_url	string
minute	int64
redistribution_allowed	bool
source	string
source_url	string
time_string	string
transformation	string
split	string
image	image

---

📊 Qualitative Results

Sample predictions and visualizations from our models on the TickTockVQA benchmark.

Effect of Swap-DPO

Qualitative examples of hand-swap error correction by Swap-DPO. SFT incorrectly swaps the hour and minute hands, whereas Swap-DPO successfully corrects this systematic error pattern.

---

📁 Files

File	Description
train-*.parquet	Training split shards with embedded images and full annotation metadata
test-*.parquet	Test split shards with embedded images and full annotation metadata
dataset_statistics.json	Dataset statistics (source distribution, splits, license distribution)
qualitative_results.png	Qualitative results (sample predictions and visualizations)
handswap_qualitative_results.png	Hand-swap error correction by Swap-DPO

The previous standalone annotations.json layout has been folded into the Parquet columns to make the dataset viewer-compatible.

---

📜 Licensing

This dataset aggregates images from multiple sources with mixed licensing terms. Do not assume a single repo-wide content license applies to every example. Use the per-row metadata fields such as source, license, license_type, license_url, commercial_use, redistribution_allowed, and attribution_required when filtering or redistributing data.

Source Distribution

Source	Rows
OpenImages	1940
coco2017	2063
ClockMovies	1244
VisualGenome	1246
CC12M	3677
ImageNet	780
SBU	1533

License Distribution

License	Rows
CC BY 4.0	8926
Unknown/Custom	1244
Custom (Non-commercial Research Use)	2313

Notes

• ampm, source_url, and license_url can be null for some examples.
• design is a list-valued field preserved from the original annotations.

---

📚 Citation

If you find our work useful, please cite:

@article{choi2026clockreasoning,
  title   = {It's Time to Get It Right: Improving Analog Clock Reading and Clock-Hand Spatial Reasoning in Vision-Language Models},
  author  = {Choi, Jaeha and Lee, Jin Won and You, Siwoo and Lee, Jangho},
  journal = {arXiv preprint arXiv:2603.08011},
  year    = {2026},
  url     = {https://arxiv.org/abs/2603.08011}
}

If you find our work useful, we would appreciate a Hugging Face 🤗 like and a GitHub ⭐ Star. Thank you!

---

🔗 Links

• GitHub: https://github.com/allchiever/It-s-Time-to-Get-It-Right
• Project Page: https://it-s-time-to-get-it-right.github.io/
• arXiv: https://arxiv.org/abs/2603.08011