[Paper Review] Optical Illusions Images Dataset

Robert Max Williams, Roman V. Yampolskiy|arXiv (Cornell University)|Sep 30, 2018

Advanced Vision and Imaging3 references8 citations

TL;DR

This paper introduces a dataset of 6,725 optical illusion images collected from two major websites, along with a curated subset of 500 hand-picked images, to advance research in computer vision and human perception. The study demonstrates that a transfer learning classifier can distinguish illusion categories better than random, but standard GANs fail to generate meaningful illusions, highlighting the need for deeper models of human visual processing.

ABSTRACT

Human vision is capable of performing many tasks not optimized for in its long evolution. Reading text and identifying artificial objects such as road signs are both tasks that mammalian brains never encountered in the wild but are very easy for us to perform. However, humans have discovered many very specific tricks that cause us to misjudge color, size, alignment and movement of what we are looking at. A better understanding of these phenomenon could reveal insights into how human perception achieves these feats. In this paper we present a dataset of 6725 illusion images gathered from two websites, and a smaller dataset of 500 hand-picked images. We will discuss the process of collecting this data, models trained on it, and the work that needs to be done to make it of value to computer vision researchers.

Motivation & Objective

• To create a comprehensive, publicly accessible dataset of static optical illusion images for use in computer vision and perception research.
• To evaluate whether existing machine learning models can recognize and generalize across different types of optical illusions.
• To investigate the limitations of current generative models, such as GANs, in learning to produce novel, human-like illusions.
• To explore whether artificial systems can be trained to perceive illusions in ways analogous to human visual perception.
• To identify gaps in current deep learning approaches for modeling human visual illusions and suggest future research directions.

Proposed method

• Collected 6,436 images from Mighty Optical Illusions and 1,454 from ViperLib using web scrapers, preserving metadata such as categories and page titles.
• Constructed a curated subset of 500 hand-picked images with strong, unambiguous visual effects for focused analysis.
• Trained a transfer learning classifier using a pretrained 'bottleneck' model fine-tuned on illusion categories, leveraging features from a larger general image dataset.
• Attempted to train a GAN using HyperGAN on a subset of the data without hyperparameter tuning, using random noise vectors as input.
• Evaluated model performance using confusion matrices and training progression visualizations to assess classification and generation quality.
• Proposed future approaches such as human-in-the-loop feedback or training on proxies for human vision to improve illusion generation.

Experimental results

Research questions

• RQ1 Can a deep learning classifier trained on general image features distinguish between different types of optical illusions?
• RQ2 To what extent can standard GAN architectures generate new, convincing optical illusions from a limited dataset of static images?
• RQ3 Why do standard generative models fail to produce meaningful illusions despite being trained on a dataset of perceptually salient images?
• RQ4 What role does human feedback or a proxy for human vision play in enabling artificial systems to learn to create or perceive illusions?
• RQ5 How do the structural and perceptual properties of optical illusions limit the applicability of current deep generative models?

Key findings

• The transfer learning classifier achieved significantly better-than-random performance in distinguishing illusion categories, indicating that visual features of illusions are detectable by deep neural networks.
• The confusion matrix for the classifier showed meaningful clustering of predictions, suggesting that the model learned to associate specific visual patterns with illusion types.
• A GAN trained on a subset of the dataset failed to generate coherent or diverse illusion-like images after 7 hours of training, producing images with low variation and no recognizable illusion effects.
• The GAN training progression revealed abnormal lack of diversity in generated outputs, suggesting mode collapse or poor optimization despite using a standard training pipeline.
• The failure of GANs to generate illusions indicates that current generative models lack an understanding of the perceptual principles underlying human visual illusions.
• The dataset's small size and limited variety of illusion types—likely only a few dozen unique categories—pose a significant challenge for training effective generative models.