clip-oscope/simco-comco

ComCo & SimCo Datasets

🔗 GitHub Project Page | 📄 arXiv Paper

Overview

This repository contains two datasets, ComCo and SimCo, designed for evaluating multi-object representation in Vision-Language Models (VLMs). These datasets provide controlled environments for analyzing model biases, object recognition, and compositionality in multi-object scenarios.

• ComCo: Composed of real-world objects derived from the COCO dataset.
• SimCo: Contains simple geometric shapes in structured multi-object settings.

---

ComCo Dataset

The ComCo (Complex COCO Objects) dataset consists of images featuring 2 to 5 objects from the COCO dataset. Each zip file contains different arrangements of objects with variations in:

• Size (e.g., large vs. small objects)
• Position (top-left, middle, bottom-right, etc.)

ComCo is specifically designed to test VLMs on real-world objects, allowing precise control over object placement and ensuring a systematic evaluation of compositional understanding.

SimCo Dataset

The SimCo (Simple Compositional Objects) dataset consists of synthetic images featuring geometric shapes such as:

• Cubes
• Spheres
• Cylinders
• Triangles
• Pentagons

SimCo is used to isolate model biases by removing real-world semantics, enabling controlled evaluation of how VLMs process object interactions purely based on size, shape, and position.

Usage

These datasets are useful for:

• Analyzing VLM biases (e.g., preference for larger objects)
• Compositionality testing (how models handle multiple objects in images)
• Zero-shot & fine-tuning tasks (evaluating robustness of vision-language embeddings)

Loading with Hugging Face datasets Library

You can load the dataset directly using:

from datasets import load_dataset

# Load ComCo dataset
comco = load_dataset("clip-oscope/simco-comco", data_dir="ComCo")

# Load SimCo dataset
simco = load_dataset("clip-oscope/simco-comco", data_dir="SimCo")

Citation

If you use this dataset in your research, please cite:

@inproceedings{abbasi2025clip,
  title={CLIP Under the Microscope: A Fine-Grained Analysis of Multi-Object Representation},
  author={Abbasi, Reza and Nazari, Ali and Sefid, Aminreza and Banayeeanzade, Mohammadali and Rohban, Mohammad Hossein and Soleymani Baghshah, Mahdieh},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2025}
}