video: Image classification is one of AI’s most common tasks, where a system is required to recognize an object from a given image. Yet real life requires us to recognize not a single standalone object but rather multiple objects appearing together in a given image. This reality raises the question: what is the best strategy to tackle multi-object classification? The common approach is to detect each object individually and then classify them. But new research challenges this customary approach to multi-object classification tasks. In an article published today in Physica A, researchers from Bar-Ilan University in Israel show how classifying objects together, through a process known as Multi-Label Classification (MLC), can surpass the common detection-based classification.
Credit: Prof. Ido Kanter, Bar-Ilan University
Image classification is one of AI’s most common tasks, where a system is required to recognize an object from a given image. Yet real life requires us to recognize not a single standalone object but rather multiple objects appearing together in a given image.
This reality raises the question: what is the best strategy to tackle multi-object classification? The common approach is to detect each object individually and then classify them. But new research challenges this customary approach to multi-object classification tasks.
In an article published today in Physica A, researchers from Bar-Ilan University in Israel show how classifying objects together, through a process known as Multi-Label Classification (MLC), can surpass the common detection-based classification.
“Detection requires recognizing each object individually and then performing the classification on each of these objects individually,” said Prof. Ido Kanter, of Bar-Ilan's Department of Physics and Gonda (Goldschmied) Multidisciplinary Brain Research Center, who led the research. “Even with assuming perfect identification, the network will need to correctly classify each object independently whereas with MLC object combinations are classified together and not separately.”
“This new method allows the network to learn correlations between objects that appear together, which makes them more recognizable,” said PhD student Ronit Gross, a key contributor to this research. “Learning combinations, rather than just single objects, can yield better results when the network is required to recognize multiple objects. This new understanding can pave the way for AI which can better recognize object combinations in a single image,” she added.
These results question the current understanding of how multiple objects are recognized and can improve real-life applications, such as autonomous vehicles which require analyzing many objects presented together at any given moment.
A video on Multilabel versus detection classification: https://www.youtube.com/watch?v=uCwtJa4ZUDE
Journal
Physica A Statistical Mechanics and its Applications
Article Title
Multilabel classification outperforms detection-based technique
Article Publication Date
10-Dec-2024