README.md

---
pretty_name: OccludeBench
language:
- en
task_categories:
- visual-question-answering
- image-to-text
size_categories:
- n<1K
tags:
- occlusion
- visual-reasoning
- shadows
- reflections
- vision-language-models
configs:
- config_name: default
  data_files:
  - split: test
    path: images/sec3_occludebench/occluded/*
---

# OccludeBench: Can VLMs Find What's Missing?

<p align="center">
  <img src="https://huggingface.co/datasets/afs07fda89sdfas90/data/resolve/main/assets/teaser.png" alt="OccludeBench teaser. Top: a direct example (answer: balance ball) and an inferential example (answer: chimney), next to each model's accuracy on direct versus inferential images; every model's gap is larger than the humans' 5.1 points. Bottom: the four OccludeTrace stages (detect cue, recognize cue, map to source, recover in context); only the last stage separates solved from failed images, 91% vs. 50%." width="100%">
</p>

OccludeBench tests whether vision-language models can recover an object hidden from view. Each of the 378 photographs has one object fully covered by a mask, and the model is asked:

> What is the primary object that is completely occluded in this image? Be as specific as you can.

In **direct** images (197) the object can only be recovered from its shadow or reflection; in **inferential** images (181) it can be inferred from the rest of the scene. Direct images are specular reflections (94), diffuse reflections (56) or shadows (47); inferential images cover technology (45), household (44), culture (37), activities (31) and nature (24).

Answers are open-ended. A response counts as correct if either of two LLM judges (GPT-5.4, Claude Opus 4.6) accepts it, which agrees with author labels on 98% of items.

## Loading

```python
from datasets import load_dataset

ds = load_dataset("afs07fda89sdfas90/data", split="test")  # occluded images + id, type, category, answer
```

Or read `annotations.json` directly and open the image paths it lists.

## Structure

```
annotations.json                                    # One record per image
images/
  sec3_occludebench/occluded/                       # 378  Benchmark input: object behind a black, yellow-bordered mask
  sec3_occludebench/original/                       # 378  Unedited photographs
  sec4_2_why_vlms_fail/seeing_failure/              # 136  Direct images with the cue region outlined
  sec4_2_why_vlms_fail/stage2_zoomed_recognition/   # 153  Cue regions cropped from direct images
  sec4_2_why_vlms_fail/stage3_source_mapping/       # 107  Four candidate regions masked and labeled A–D
  appendix/box_sensitivity/                         # 378  Blue-filled mask, for the mask-style check
code/                                               # Evaluation, grading and analysis code
```

Images are named by item `id` (e.g. `pic0001.png`); crops add a suffix (`pic0002-1.png`).

## Annotations

| Field | Description |
|---|---|
| `id` | Image id, `pic0001`–`pic0378` |
| `type` | `direct` or `inferential` |
| `category` | Direct: `specular reflection`, `diffuse reflection`, `shadow`. Inferential: `technology`, `household`, `culture`, `activities`, `nature` |
| `fine_category` | Finer annotation label the category was merged from |
| `answer` | Reference answer |
| `occluders` | Mask geometry in original-image pixels: `bbox` (`x1`, `y1`, `x2`, `y2`) or `polygon` (`points`) |
| `images` | Paths to this item's image in each folder (`null` where a variant does not exist) |
| `cue_crops` | Crops of the shadow/reflection region: `image` path and crop `box` (`x`, `y`, `w`, `h`) in original-image pixels. Empty for items without a crop |
| `reverse_mapping` | Source-mapping `image` and the `answer` letter of the region hiding the object, or `null` |
| `source_url` | Where the photograph was found |
| `annotator_note` | Annotator's rationale for the answer |

## Reproducing the experiments

```bash
cd code
pip install -r requirements.txt
cp .env.example .env      # API keys
python prepare_data.py    # lays out the images the way the scripts expect
```

See `code/README.md` for the main commands and the code layout.

## Notes

- `pic0360`: its occluded image was made from a differently cropped copy of the photograph, so its `occluders` coordinates match its original and box-sensitivity images but not its occluded image.
- Images come from the web, Visual Genome, and the authors' own photographs. Web images remain under their original owners' terms; see `source_url`.

## Citation

```bibtex
@misc{cekinmez2026occludebench,
  title={OccludeBench: Can VLMs Find What's Missing?},
  author={Cekinmez, Jasin and Wu, Addison J. and Mitsuhashi, Ryo and Cai, Linrong and Fu, Xingyu and Liu, Zhuang},
  year={2026}
}
```