Checking the Neural Networks' Subitizing Ability and Proposing Three Challenges for Human-Like Subitizing

Subitizing refers to the ability to instantly and accurately determine the number of objects in a visual scene without counting them individually. Research in neuroscience and cognitive science suggests that humans can typically subitize up to four objects. What is remarkable about human vision intelligence is not just the ability to subitize, but also the capacity to do so with objects that have never been seen before. In this paper, we conduct a series of experiments to verify whether existing neural networks (MLP, CNN, ViT) can perform human-like subitizing. The experimental results indicate that, for objects encountered during the training process, MLPs achieved an average accuracy of 97.5%, CNNs 99%, and ViTs 99%. However, for datasets containing different geometrical patterns that were not seen during training, the accuracy of these models was significantly lower. These findings lead us to conclude that conventional training methods used in classification tasks are insufficient for achieving human-like subitizing. In this paper, we introduce three key challenges(Meta Learning Strategy for Unseen Objects Subitizing, Query Ojbect Subitizing, Knowledge Mixing based Unsupervised Query Object Subitizing) that need to be addressed for achieving human-like subitizing and propose a direction for future research in this domain.