Challenge the ELVIS by yourself: Click Here

Category Example: Click Here

Citation

@inproceedings{sha2025gestalt,
  author    = {Jingyuan Sha and Hikaru Shindo and Kristian Kersting and Devendra Singh Dhami},
  title     = {Gestalt Vision: A Dataset for Evaluating Gestalt Principles in Visual Perception},
  booktitle = {Proceedings of the 19th International Conference on Neurosymbolic Learning and Reasoning (NeSy)},
  year      = {2025}
}

Gestalt Pattern Reasoning Benchmark

Overview

This repository contains a dataset and benchmarking framework for Gestalt pattern reasoning. The dataset consists of thousands of procedurally generated visual patterns based on Gestalt principles, including proximity, similarity, closure, symmetry, and continuity. The benchmark is designed to evaluate both human and AI performance in recognizing and reasoning about these patterns.

Dataset Loading

The dataset has already been generated and uploaded to the Huggingface. This repository on the Github also includes all the code for generating the dataset.

Installation

To use this benchmark, first clone the repository and install dependencies:

git clone https://github.com/ml-research/ELVIS.git
cd ELVIS
pip install -r requirements.txt

Pattern Generation

To generate all patterns across all principles, run:

python scripts/main.py

Generated patterns will be saved in data/raw_patterns/.

Benchmarking AI Models

To evaluate AI models on the dataset:

python scripts/evaluate_models.py 

Results will be saved in data/results/.

Gestalt Principles and Patterns

The benchmark includes five Gestalt principles:

• Proximity
• Similarity
• Closure
• Symmetry
• Continuity

For each principle, there more than 100 tasks. Each task includes:

• x positive images and x negative images for training.
• x positive images and x negative images for testing.
• (set x value in config.py)

Patterns are generated using basic objects such as:

• Triangle
• Circle
• Square

Each pattern has its own folder within the respective principle directory, containing train and test subdirectories, and each pattern is named in the format id_pattern_name.

Using Docker

To ensure compatibility across environments, you can use Docker to build and run the project.

Docker Usage

git clone https://github.com/ml-research/ELVIS.git

Build

docker build -t elvis .

Run

docker run -it --gpus all -v /home/ml-jsha/ELVIS:/app -v /home/ml-jsha/storage/ELVIS_RESULT/:/elvis_result -v /home/ml-jsha/storage/ELVIS_Data/:/gen_data --rm elvis:latest 

ln -s /home/ml-jsha/nesy_causal_p/storage/dataset/grb /home/ml-jsha/ELVIS/grb

python -m scripts.main --remote

train llama

python -m scripts.evaluate_models --model llama --batch_size 2 --principle proximity --img_num 3 --device_id 0

train vit

python -m scripts.evaluate_models --batch_size 100 --principle proximity --img_num 3 --model vit --device_id 1 python -m scripts.evaluate_models --batch_size 100 --principle similarity --img_num 3 --model vit --device_id 1 python -m scripts.evaluate_models --batch_size 100 --principle closure --img_num 3 --model vit --device_id 2 python -m scripts.evaluate_models --batch_size 100 --principle symmetry --img_num 3 --model vit --device_id 3 python -m scripts.evaluate_models --batch_size 100 --principle continuity --img_num 100 --device_id 5

train Llava

python -m scripts.evaluate_models --batch_size 1 --principle proximity --img_num 3 --model llava --device_id 0 python -m scripts.evaluate_models --batch_size 1 --principle similarity --img_num 3 --model llava --device_id 2 python -m scripts.evaluate_models --batch_size 1 --principle closure --img_num 3 --model llava --device_id 3 python -m scripts.evaluate_models --batch_size 1 --principle symmetry --img_num 3 --model llava --device_id 4 python -m scripts.evaluate_models --batch_size 100 --principle continuity --img_num 3 --model llava --device_id 7

train deepseek

python -m scripts.evaluate_models --batch_size 1 --principle proximity --model deepseek --device_id 2

## File Structure

gestalt_benchmark/ │── data/ │ │── raw_patterns/ # Unprocessed/generated raw patterns │ │ │── proximity/ │ │ │ │── train/ │ │ │ │ │── 0001_red_triangle/ │ │ │ │ │ │── positive/ │ │ │ │ │ │ │── 00000.png │ │ │ │ │ │ │── 00000.json │ │ │ │ │ │ │── 00001.png │ │ │ │ │ │ │── 00001.json │ │ │ │ │ │ │── 00002.png │ │ │ │ │ │ │── 00002.json │ │ │ │ │ │── negative/ │ │ │ │ │ │ │── 00000.png │ │ │ │ │ │ │── 00000.json │ │ │ │ │ │ │── 00001.png │ │ │ │ │ │ │── 00001.json │ │ │ │ │ │ │── 00002.png │ │ │ │ │ │ │── 00002.json │ │ │ │── test/ │ │ │ │ │── 0001_red_triangle/ │ │ │ │ │ │── positive/ │ │ │ │ │ │ │── 00000.png │ │ │ │ │ │ │── 00000.json │ │ │ │ │ │ │── 00001.png │ │ │ │ │ │ │── 00001.json │ │ │ │ │ │ │── 00002.png │ │ │ │ │ │ │── 00002.json │ │ │ │ │ │── negative/ │ │ │ │ │ │ │── 00000.png │ │ │ │ │ │ │── 00000.json │ │ │ │ │ │ │── 00001.png │ │ │ │ │ │ │── 00001.json │ │ │ │ │ │ │── 00002.png │ │ │ │ │ │ │── 00002.json │ │ │── similarity/ │ │ │ │── train/ │ │ │ │── test/ │ │ │── closure/ │ │ │ │── train/ │ │ │ │── test/ │ │ │── symmetry/ │ │ │ │── train/ │ │ │ │── test/ │ │ │── continuity/ │ │ │ │── train/ │ │ │ │── test/ │── scripts/ │ │── proximity/ # Folder containing scripts for proximity patterns │ │ │── prox_patterns.py │ │ │── util... │ │── similarity/ │ │── closure/ │ │── symmetry/ │ │── continuity/ │ │── utils/ # Folder containing utility functions │ │── main.py # Entry point to generate all patterns of all principles │ │── config.py # Configuration settings for colors, shapes, and numbers │ │── evaluate_models.py # Benchmarking models │── README.md │── requirements.txt

## Contribution
We welcome contributions to improve the dataset and evaluation framework. Please submit pull requests with explanations of changes.

## License
This project is licensed under the MIT License.

---
🚀 **Ready to challenge AI with Gestalt patterns? Start now!**