OPTICS: Commonsense Reasoning over Visual Object Properties

Inspired by human object categorization, OPTICS (formerly ORBIT) is a benchmark suite for evaluating how Vision–Language Models (VLMs) perceive and reason about object properties. It spans four object property dimensions (physical, taxonomic, functional, relational), three levels of reasoning complexity (direct recognition, property inference, counterfactual reasoning), and three image types (photographic/real, animated, AI-generated). It has two parts: OPTICS-CNT (exact counting) and OPTICS-CMP (count comparison).

📄 Paper | 📚 arXiv | 💻 Code

Uses

This dataset tests VLMs with count-based questions to evaluate how well they reason about object properties across visual contexts. Load each part as its own config:

from datasets import load_dataset

cnt = load_dataset("Abk802/OPTICS", "OPTICS-CNT", split="train")
cmp = load_dataset("Abk802/OPTICS", "OPTICS-CMP", split="train")

Repository structure

OPTICS-CNT/
├── data/        # images + metadata.csv per image type
└── <benchmark>.json
OPTICS-CMP/
├── merged_images/        # merged image pairs + metadata.csv
└── <merged_benchmark>.json

OPTICS-CNT

360 images (120 per image type) with 1,080 counting questions, 3 per image, one per reasoning level. Ground-truth counts range from 0 to 10.

Fields:

• image_id: the image number (120 under each image_type).
• image_type: REAL, ANIMATED or AI_GENERATED.
• theme: label indicating the image theme.
• reasoning_complexity: direct_recognition, property_inference or counterfactual_reasoning.
• question_id: the question number (one per reasoning level).
• question: the counting question.
• answer: the ground-truth count.
• property_category: physical, taxonomic, functional or relational.
• objects: annotator notes on the counted objects (reasoning for answer; not given to models).

OPTICS-CMP

2,122 comparison questions built by pairing semantically similar OPTICS-CNT questions (cosine similarity ≥ 0.80). Each example is a left/right concatenation of two images, and the model decides which side has the larger count. Equal-count pairs are excluded, and count differences range from 1 to 9.

Fields:

• question: the comparison question ("Let A be the answer to … for the left image. Let B be the answer to … for the right image. Which is greater, A or B?").
• answer: A or B (evaluation label).
• answer_side: LEFT or RIGHT.
• property_category, reasoning_complexity: category of the paired questions.
• count_diff: absolute difference between the two ground-truth counts.
• question1, image_id1, image_type1, count1: left-hand question, image and count.
• question2, image_id2, image_type2, count2: right-hand question, image and count.
• similarity: cosine similarity of the paired question embeddings.

Licensing

• Code and annotations (questions and answers): MIT License.
• Images remain subject to their source licenses:
  • Unsplash: Unsplash License (free to use, modify and redistribute, no attribution required).
  • Freepik: Freepik Free License (free tier, attribution required).
  • Google Images: used only as a search interface. Most images were traced to Unsplash or Freepik. Fewer than 10 untraceable images are used solely for non-commercial research benchmarking.
  • AI-generated images (GPT-4o, Grok 3): generated by the authors, subject to OpenAI and xAI usage terms.

Citation

@misc{kolari2026studycommonsensereasoningvisual,
      title={A Study of Commonsense Reasoning over Visual Object Properties}, 
      author={Abhishek Kolari and Mohammadhossein Khojasteh and Yifan Jiang and Floris den Hengst and Filip Ilievski},
      year={2026},
      eprint={2508.10956},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.10956}, 
}