NetanelTamir/Stuck-On-Shuffle-SABER

SABER: Synthetic Attribute Binding Evaluation for Representations

Project Page  |  arXiv  |  Code  |  Finetuning Dataset

A benchmark for attribute-object binding in vision-language models, covering all four text and image modality pairings (text–text, text–image, image–text, and image–image). Scenes contain 2 to 10 objects, built from 232 everyday objects and 50 visual attributes. Each of the 1,800 multiple-choice questions has a query and four candidates, all given both as an image and as a caption; the candidates contain exactly the same objects and attributes as the query, and only one keeps each attribute on the right object. The benchmark is structured as multiple-choice questions, but the candidates can also be pooled into a single gallery for large-scale retrieval evaluation.

Files

file	rows	content
2_objects.csv ... 10_objects.csv	200 each	one file per number of objects in the scene, one question per row
images/<n>_objects/*.png	9,000	the rendered scenes, 1024 x 1024, named <id>_query.png and <id>_ans_<k>.png

Columns

column	content	example (2_objects.csv, row 0)
id	row index within the file	0
query_img	repo-relative path of the query image	images/2_objects/0_query.png
query_txt	query caption	A sandy snowboard and a pink cherry.
ans_k_img (k = 1..4)	candidate image paths	images/2_objects/0_ans_1.png
ans_k_txt (k = 1..4)	candidate captions	A pink cherry and a sandy snowboard.

Candidates

candidate	construction
ans_1	correct: the query's bindings, with the pairs in a different order
ans_2	attributes permuted across the objects
ans_3	objects permuted across the attributes
ans_4	both permuted (for 2 objects, where only two assignments exist: one pair replaced by a fresh object or attribute)

Per object count, rows 0 to 99 are sampled independently and rows 100 to 199 form nested chains: the N-object query contains the (N-1)-object query of the same index plus one more pair.

Evaluation

SABER supports two complementary protocols. Because every item exists as both an image and a caption, each protocol runs in all four modality pairings (text–text, text–image, image–text, and image–image).

Multiple choice. For each question, we score the similarity between the query and each of its four candidates and select the highest-scoring candidate. The prediction is correct if ans_1 ranks first. Because all four candidates contain the same objects and attributes and differ only in their attribute–object assignments, success requires distinguishing the bindings. In our paper we report accuracy for each setting as the unweighted mean across the nine object counts; chance performance is 25%.

Large-gallery retrieval. We pool the candidates from all questions into a single gallery of 7,200 items (four candidates for each of 1,800 questions across all object counts). Each of the 1,800 queries then ranks the full gallery by similarity. A gallery item is considered positive if it contains all attribute–object bindings in the query, even if it includes additional bindings. Items containing the same objects and attributes but different bindings serve as hard negatives. In our paper we report mean average precision (mAP), separately for the two halves of each CSV:

subset	rows	positives
independent	id 0 to 99	exactly one: the question's own ans_1
nested	id 100 to 199	the question's own ans_1, plus the ans_1 items of the same chain at larger object counts, whose scenes contain the query's bindings inside a bigger scene

The independent subset measures whether a query's bindings are recovered against thousands of distractors that share vocabulary with it. The nested subset additionally tests whether the queried bindings are recognised when they appear inside a larger scene, rather than relying on overlapping objects and attributes alone.

Citation

The paper is under review; a citation entry will be added when it is public.