README.md

---
annotations_creators:
- expert-generated
language:
- en
license: cc-by-sa-4.0
size_categories:
- 1K<n<10K
task_categories:
- visual-question-answering
- image-text-to-text
task_ids: []
pretty_name: VisualOverload
tags:
- fiftyone
- image
- vqa
- visual-question-answering
- art
dataset_summary: |



  ![image/png](dataset_preview.jpg)


  This is a [FiftyOne](https://github.com/voxel51/fiftyone) dataset with 2720 samples,
  converted to FiftyOne format from the original
  [paulgavrikov/visualoverload](https://huggingface.co/datasets/paulgavrikov/visualoverload).


  ## Installation


  If you haven't already, install FiftyOne:


  ```bash
  pip install -U fiftyone
  ```


  ## Usage


  ```python
  import fiftyone as fo
  from fiftyone.utils.huggingface import load_from_hub

  # Load the dataset
  # Note: other available arguments include 'max_samples', etc
  dataset = load_from_hub("Voxel51/VisualOverload")

  # Launch the App
  session = fo.launch_app(dataset)
  ```
---

# Dataset Card for VisualOverload

![image/png](dataset_preview.jpg)

This is a [FiftyOne](https://github.com/voxel51/fiftyone) dataset with **2,720 samples**.
It is a FiftyOne-format conversion of the original
[**paulgavrikov/visualoverload**](https://huggingface.co/datasets/paulgavrikov/visualoverload)
dataset (CVPR 2026). All credit for the data, annotations, and benchmark design belongs to
the original authors — please see [Citation](#citation) and
[Dataset Sources](#dataset-sources).

## Installation

If you haven't already, install FiftyOne:

```bash
pip install -U fiftyone
```

## Usage

```python
import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/VisualOverload")

# Launch the App
session = fo.launch_app(dataset)
```

## Dataset Details

### Dataset Description

Is basic visual understanding really solved in state-of-the-art VLMs? **VisualOverload** is a
visual question answering (VQA) benchmark comprising **2,720 question–answer pairs** with
privately held ground-truth responses. Unlike prior VQA datasets that typically focus on
near-global image understanding, VisualOverload challenges models to perform simple,
knowledge-free vision tasks in densely populated (or *overloaded*) scenes. The dataset
consists of **150 high-resolution scans of public-domain paintings** populated with multiple
figures, actions, and unfolding subplots set against elaborately detailed backdrops. The
images were manually annotated with questions across six task categories to probe a thorough
understanding of the scene.

The authors hypothesize that current benchmarks overestimate the performance of VLMs, and
that encoding and reasoning over details remains challenging, especially in densely populated
scenes. Indeed, even the best model evaluated (o3) out of 37 tested models reaches only
**19.6% accuracy on the hardest split** and **69.5% overall**. The accompanying error
analysis reveals failure modes including weak counting, OCR failures, and logical
inconsistencies under complex tasks.

- **Curated by:** Paul Gavrikov, Wei Lin, M. Jehanzeb Mirza, Soumya Jahagirdar, Muhammad Huzaifa, Sivan Doveh, Serena Yeung-Levy, James Glass, and Hilde Kuehne
- **Shared by:** [Voxel51](https://huggingface.co/Voxel51) (FiftyOne-format conversion)
- **Language(s) (NLP):** en
- **License:** CC BY-SA 4.0 (the underlying images are royalty-free public-domain artwork, CC0)

### Dataset Sources

- **Original dataset (please cite this):** https://huggingface.co/datasets/paulgavrikov/visualoverload
- **Repository:** https://github.com/paulgavrikov/visualoverload
- **Paper:** [VisualOverload: Probing Visual Understanding of VLMs in Really Dense Scenes (arXiv:2509.25339)](https://arxiv.org/abs/2509.25339)
- **Project page:** https://paulgavrikov.github.io/visualoverload/
- **Leaderboard / online evaluator:** https://huggingface.co/spaces/paulgavrikov/visualoverload-submit

## Uses

### Direct Use

- Benchmark the fine-grained visual understanding of vision-language models (VLMs) in dense,
  detail-heavy scenes.
- Slice and analyze results by **question type**, **difficulty**, and **category** using the
  prefixed sample tags (see [Dataset Structure](#dataset-structure)).
- Run a VLM per question — each sample carries a single `question` (and a ready-to-use
  `default_prompt`), so a model can read the prompt from the sample field and write one
  prediction per sample, then submit `question_id` + predicted answer to the official
  evaluation server.

### Out-of-Scope Use

- **Training / fine-tuning.** This is an evaluation benchmark; ground-truth answers are held
  privately and are intentionally not distributed.
- Drawing conclusions about general image understanding outside the dense-scene,
  painting-domain setting the benchmark was designed for.

## Dataset Structure

The benchmark is modeled **one sample per question**: **2,720 samples** over **150 paintings**
(each image is shared by the ~18 questions that reference it). Ground-truth answers are **not
included** — models are scored via the official evaluation server using each question's
`question_id`. All samples belong to the single `test` split.

**Fields**

| Field | Type | Description |
|-------|------|-------------|
| `filepath` | image | Path to the painting (shared across its questions) |
| `question_id` | `StringField` | Unique id — the key used for leaderboard submissions |
| `question` | `StringField` | The question about the image |
| `response_options` | `ListField(StringField)` | Answer options for `choice` questions (e.g. `["yes", "no"]`); empty otherwise. Listed as `options` in the source dataset. |
| `default_prompt` | `StringField` | Ready-to-use prompt (question + options + output-format constraint) |
| `image_id` | `StringField` | Painting id (filename stem) — groups an image's questions |
| `win_rate` | `FloatField` | Per-image model win-rate from the benchmark (a difficulty signal) |
| `metadata` | `ImageMetadata` | Image width/height (most images are ~4K, e.g. 3840×2160) |

**Sample tags** — `question_type`, `difficulty`, and `category` are stored as **prefixed
sample tags** (filter via the App sidebar or `dataset.match_tags(...)`). They are prefixed
because `question_type` and `category` share the values `counting` and `ocr`.

| Tag prefix | Values (counts) |
|------------|-----------------|
| `question_type:` | `choice` (2043), `counting` (559), `ocr` (118) |
| `difficulty:` | `easy` (986), `medium` (1304), `hard` (430) |
| `category:` | `activity` (150), `attributes` (149), `counting` (559), `ocr` (118), `reasoning` (356), `scene` (1388) |

Every sample is also tagged `test`.

```python
# Example: all hard OCR questions
from fiftyone import ViewField as F
hard_ocr = dataset.match_tags(["difficulty:hard", "question_type:ocr"], all=True)
```

## Dataset Creation

### Curation Rationale

Existing VQA benchmarks largely probe near-global image understanding and may overestimate
VLM capability. VisualOverload deliberately targets *simple, knowledge-free* perception
(reading, counting, attribute and activity recognition, scene/relationship reasoning) in
**overloaded** scenes that contain many figures, actions, and subplots, to expose the gap in
encoding and reasoning over fine detail.

### Source Data

#### Data Collection and Processing

The images are high-resolution scans of **public-domain paintings** (CC0). Most match a 4K
pixel budget (≈ 3840×2160) across varying aspect ratios.

#### Annotations

The images were **manually annotated** with questions spanning six task categories
(`activity`, `attributes`, `counting`, `ocr`, `reasoning`, `scene`), three difficulty levels
(`easy`, `medium`, `hard`), and three answer/question types (`choice` with 2 or 4 options,
freeform `counting`, and freeform `ocr`). Ground-truth answers are withheld to prevent
contamination and are only accessible through the evaluation server.

## Bias, Risks, and Limitations

- **Evaluation-only:** ground truth is private; scoring requires the official server, so this
  copy cannot be used for supervised training or offline scoring.
- **Domain:** the imagery is limited to scanned public-domain paintings; performance here may
  not transfer to photographs or other domains.
- **Scale:** 150 source images / 2,720 questions — small relative to large-scale VQA corpora.

### Recommendations

Use VisualOverload as a targeted probe of fine-grained perception in dense scenes rather than
a general VQA score. Report results by difficulty and category (the sample tags make this
easy) and submit predictions to the official evaluator for comparable, leak-free numbers.

## Citation

If you use this dataset, please cite the original work:

**BibTeX:**

```bibtex
@InProceedings{Gavrikov_2026_visualoverload,
  author    = {Paul Gavrikov and Wei Lin and M. Jehanzeb Mirza and Soumya Jahagirdar and Muhammad Huzaifa and Sivan Doveh and Serena Yeung-Levy and James Glass and Hilde Kuehne},
  title     = {{VisualOverload}: Probing Visual Understanding of VLMs in Really Dense Scenes},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  month     = {June},
  year      = {2026}
}
```

**APA:**

Gavrikov, P., Lin, W., Mirza, M. J., Jahagirdar, S., Huzaifa, M., Doveh, S., Yeung-Levy, S., Glass, J., & Kuehne, H. (2026). *VisualOverload: Probing Visual Understanding of VLMs in Really Dense Scenes.* In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).

## Dataset Card Authors

FiftyOne-format conversion shared by [Voxel51](https://huggingface.co/Voxel51). The dataset,
annotations, and benchmark were created by Paul Gavrikov et al.; see the original dataset at
[paulgavrikov/visualoverload](https://huggingface.co/datasets/paulgavrikov/visualoverload).