afs07fda89sdfas90/data

OccludeBench: Can VLMs See the Hint?

OccludeBench tests whether vision-language models use the specific visual evidence a question requires. Each of the 378 real photographs has one object fully hidden behind a drawn mask, and the model is asked to name it:

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

• Direct (197 images): the hidden object can only be recovered from a single visual cue, its shadow or its reflection.
• Inferential (181 images): the hidden object can be inferred from the surrounding scene.

Answers are open-ended and graded by an LLM judge against the reference answer.

Contents

Path	Images	Paper	Used for
images/sec3_occludebench/occluded/	378	§3–4	Benchmark input: target object hidden by a black mask with a yellow border
images/sec3_occludebench/original/	378	§3	Unedited photographs
images/sec4_2_why_vlms_fail/seeing_failure/	136	§4.2	Direct images with the shadow/reflection region outlined (projection-region probe)
images/sec4_2_why_vlms_fail/stage2_zoomed_recognition/	153	§4.2	Shadow/reflection regions cropped from direct images (zoomed-in recognition)
images/sec4_2_why_vlms_fail/stage3_source_mapping/	107	§4.2	Four candidate regions blacked out and labeled A–D, shadow/reflection visible (source mapping)
images/appendix/box_sensitivity/	378	App. C	Mask filled blue with a yellow border (sensitivity to mask style)
annotations.json			One record per image (fields below)
code/			Evaluation, judging and analysis code for every experiment in the paper

Folders follow the paper's sections, matching code/. Every image is named by its item id, e.g. images/sec3_occludebench/occluded/pic0001.png; crops add a suffix (pic0002-1.png).

Annotations

annotations.json is a list of 378 records:

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

Type	Category	Images
Direct	Specular reflection	94
Direct	Diffuse reflection	56
Direct	Shadow	47
Inferential	Technology	45
Inferential	Household	44
Inferential	Culture	37
Inferential	Activities	31
Inferential	Nature	24

Loading

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.

Reproducing the experiments

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

code/README.md maps each section, table and figure of the paper to the script that produces it.

Grading

A response counts as correct if either of two judges (Claude Opus 4.6, GPT-5.4) accepts it as the same object as the reference answer. This ensemble agrees with author labels on 98% of items. The judge prompt is in code/sec3_occludebench/judger_script.py.

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.

Sources

Images come from the web, Visual Genome, and the authors' own photographs. Web images remain under their original owners' terms; see source_url.