MarvinRichie/perspective-taking-bench

Perspective-Taking Benchmark

A synthetic VLM benchmark for spatial perspective-taking. Each example is a rendered scene containing subjects (people) and everyday objects, paired with a question that requires reasoning about spatial relationships from a specific subject's point of view rather than the camera's.

Tasks

Task	Question	Answer
spatial_lr	Is object A to the left or right of object B, from subject X's perspective?	left / right
spatial_nf	Is object A closer or farther than object B, from subject X's perspective?	closer / farther
facing	Which object/subject is subject X facing (or none of them)?	object/subject name / None
visibility_oov	Is object A visible from subject X's perspective?	yes / no
visibility_occ	Is object A visible or blocked (occluded by another object) from subject X's perspective?	yes / no

Each task has multiple settings varying the number of subjects (1 or 2) and objects (1-3) in the scene, e.g. 1subj_2obj, 2subj_3obj.

Dataset composition

3,829 rows total.

Task	Setting	Rows
facing	1subj_2obj	101
facing	1subj_3obj	106
facing	2subj_2obj	220
facing	2subj_3obj	102
spatial_lr	1subj_2obj	220
spatial_lr	1subj_3obj	220
spatial_lr	2subj_2obj	440
spatial_lr	2subj_3obj	220
spatial_nf	1subj_2obj	220
spatial_nf	1subj_3obj	220
spatial_nf	2subj_2obj	440
spatial_nf	2subj_3obj	220
visibility_occ	1subj_2obj	110
visibility_occ	1subj_3obj	110
visibility_occ	2subj_2obj	220
visibility_occ	2subj_3obj	110
visibility_oov	1subj_1obj	110
visibility_oov	1subj_2obj	110
visibility_oov	2subj_1obj	110
visibility_oov	2subj_2obj	220

Every 2subj_2obj setting is roughly double-sized because it includes a background-augmented variant of every example (see below) -- 3,059 rows use the original rendered background, 770 use an augmented one.

Schema

Each row:

Column	Type	Description
file_name	string	Relative path to the image (for imagefolder auto-loading)
task	string	One of the 5 tasks above
dataset	string	single (1 subject) or multi (2 subjects)
setting	string	e.g. 2subj_3obj
scene_id	string	Source scene identifier
question	string	The question text
answer	string	The ground-truth answer
subject_descriptor	string	The subject the question is asked from the perspective of (e.g. "man", "woman")
spatial_object_a / spatial_object_b	string or null	The two compared objects (spatial_lr/spatial_nf only)
visibility_object	string or null	The object in question (visibility_oov/visibility_occ only)
facing_object_options	list or null	The candidate objects/subjects offered as answer choices (facing only)
image_variant	string	original or background
background_setting	string or null	The place description used for the augmented background (background rows only), e.g. "cozy bedroom"

Background augmentation

For the 2subj_2obj setting of every task, one additional image variant was generated per example by editing the original render with an AI image model (OpenAI's image-edit API): the subjects/objects are preserved exactly via a segmentation mask, and only the background/floor is regenerated into a different setting (e.g. "warehouse", "beach", "art gallery"). These image_variant == "background" images are AI-generated/edited content, disclosed here for transparency. The background_setting column records which place description was used for each one.

Data generation

Scenes are synthetic, rendered in Blender from 3D assets (subjects and objects) placed in a simple room and photographed from both an overhead multi-angle "scene" camera and first-person (POV) cameras per subject. Questions and ground-truth answers are derived programmatically from each scene's known geometry (positions, facing directions, camera frustums), not by manual annotation.

Licensing

This dataset is released under CC BY-NC 4.0 (Attribution-NonCommercial). It may be used and adapted for any non-commercial purpose (including academic research) with attribution; commercial use requires separate permission. This restriction is inherited from a small number of the underlying 3D assets below, which are themselves licensed CC Attribution-NonCommercial.

Some images additionally contain AI-generated/edited content (see "Background augmentation" above).

3D asset attribution

Every scene is composed from the following 3D assets (all sourced from Sketchfab). Object/subject placement, materials, and scene composition are original to this project; the underlying 3D models are third-party works, attributed here per their license terms. Entries marked NC are licensed CC Attribution-NonCommercial; all others are CC Attribution.

Category	License	Title	Author	Source
basketball	CC BY	Standart Basketball-Ball	Umut Calim	link
cactus	CC BY	Cactus	Victoria Reinhold	link
chair	CC BY	Red Pleather Rolling Office Chair	zhixson	link
chair	CC BY	Кресло Racer (13230)	Tetchair	link
chair	CC BY	Кресло Swan (15333)	Tetchair	link
chalkboard	CC BY	[FREE] Restaurant Black Chalkboard Blackboard 4K	Rodrigo E. Maranzana	link
closet	CC BY	Animated wardobe closet office low poly	SusanKing	link
closet	CC BY	Closet - High Poly 3D Model	Ezecrafting_YT	link
closet	CC BY	Closet from Poly by Google	IronEqual	link
closet	CC BY	Low Poly closet with books and plants	Dokazvo	link
fan	CC BY	Fan (Turbo Air - High Speed Fan)	minitakesstudio	link
fan	CC BY	Stand Fan	Flighty Dreams	link
fire_hydrant	CC BY	Fire Hydrand	M3D	link
fire_hydrant	CC BY	Fire_Hydrant	yura.bruika1	link
fridge	CC BY-NC	Haier Refrigerator	cgwings	link
fridge	CC BY	Japan Vending Machine Soft Drink	ditovirnantio	link
guitar	CC BY-NC	Last of Us 2 Inspired Guitar	TierneyV3	link
lamp	CC BY-NC	Chinese lamp	Coffeek	link
lamp	CC BY	Floor Lamp	NoMoreFeelings	link
lamp	CC BY	Mid Century Modern Floor Lamp	thebiz	link
luggage	CC BY	Suitcase	moh1t	link
luggage	CC BY	Week 15 - Luggage Bake V2	KoteTheCat	link
mailbox	CC BY	Home US Mailbox	Glowbox 3D	link
mailbox	CC BY	Week 9.5 - Mailbox V2	KoteTheCat	link
shelf	CC BY	Nightstand / Тумбочка / Прикроватная тумбочка	Every 3D	link
traffic_cone	CC BY	Signal Cone	gabrielmendesm	link
traffic_cone	CC BY	Traffic Cone - Game Ready	unknown	link
trash_can	CC BY	Paper Waste Bin	ellieharrodine	link
trash_can	CC BY	Trash Can 2	BananaBagal	link
tv	CC BY	TV & Glass Table Sketchfab	James Rogers	link
tv	CC BY	Tv Unit	tdh082	link
vacuum	CC BY	Upright Dyson Vacuum Cleaner	rhcreations	link
man (subject)	CC BY	Andrew - LOD Man character	Egunoff	link
man (subject)	CC BY	Kulachet3D Avatar	kulachet.l	link
man (subject)	CC BY	My Avatar	Damiano	link
woman (subject)	CC BY	C5f74cfd-cf5c-479a-8204-5d50e4f36542 (10)	yhomi	link
woman (subject)	CC BY	Free 018 Kana Walk	ddd	link
woman (subject)	CC BY	Woman Standing Idle_Kinstyuri	yuriannoue	link

This is the complete set of 3D assets available to the scene generator -- every rendered scene draws exclusively from this pool (verified directly against the generated scene data), so this list fully covers every asset that could appear in any image in this dataset.

Citation

TODO: add citation once available.