Why humans are better at recognizing objects from fragments while AI is struggling

A study from EPFL reveals why humans are superior at recognizing objects from fragments, which underscores the key role of contour integration in human vision.

Every day we can easily recognize friends in the crowd, and identify familiar shapes, even if they are partially hidden. Our brains gather fragments all over the object, fill in blanks, and understand the often-confused world.

This ability is called “contour integration,” something even the smartest AI systems find difficult. Despite the notable achievements of artificial intelligence in image recognition, AIS still struggles to generalize from incomplete or broken visual information.

When an object is partially hidden, erased, or split into fragments, most AI models become agitated, misclassified, or given up. This can become a serious problem in real life given the increasing reliance on AI for real applications such as self-driving cars, prosthetics, and robotics.

The EPFL NeuroAI Lab, led by Martin Schrimpf, set out to systematically compare how people and AI handle visual puzzles. Edne’s graduate student and lead author of the study, Ben Lönnqvist, collaborated with Michael Herzog’s Institute of Psychophysics, developed a series of recognition tests in which both humans and over 1,000 artificial neural networks had to be identified to objects with missing or fragmented contours. Their results show that humans always outperform cutting edge AI when it comes to contour integration.

The study was presented at the International Conference on Machine Learning (ICML 2025) held in Vancouver from July 13th to 19th. It can be used with ARXIV preprint server.

The team set up a lab-based object recognition test that included 50 volunteers. Participants displayed images of everyday items such as cups, hats and frying pans. Sometimes, only 35% of the object’s outline can now be visible. In parallel, the team gave the same tasks to over 1,000 AI models, including the most powerful systems available.

This experiment covered 20 different conditions that vary the type and amount of visual information. The team compared performance in these conditions, measured accuracy, and analyzed how both humans and machines responded to increasingly difficult visual puzzles.

Humans proved to be very robust and often achieved 50% accuracy even when most of the object’s outlines were missing. In contrast, AI models tended to collapse into random inferences under the same circumstances. Only models trained with billions of images approached human-like performance. Still, it had to be particularly adapted to the images of the study.

Digging deeper, the researchers found that the teams show that humans naturally prefer to recognize objects when fragmented parts, which they call “integration bias,” point in the same direction. AI models trained to develop similar biases improved performance when challenging image distortion. Training AI systems specifically designed to integrate contours increased their accuracy and focused more on the shape of the object rather than on the surface texture.

These results suggest that contour integration can be learned from experience rather than hardwired properties. Industry relying on computer vision, such as autonomous vehicles and medical imaging, can build AI that views the world as a safer and more reliable technology.

The work also shows that the best way to close the gap is not by tinkering with the AI architecture, but by giving it a more “human-like” visual diet, containing multiple real-world images, where objects are often partially hidden.