Do deep learning models recognize 3D shapes in the same way humans do? (www.santafe.edu)

A recent study led by SFI Program Postdoctoral Fellow Shuhao Fu investigates whether deep learning models recognize 3D shapes similarly to humans. Researchers compared human observers’ performance with that of two prominent models for point-cloud recognition: the Dynamic Graph Convolutional Neural Network (DGCNN) and the transformer-based Point Transformer. The study demonstrated that while humans maintained strong recognition abilities even when point clouds were sparse or distorted, their performance significantly declined when the configuration of object parts was scrambled. This finding suggests that human 3D vision heavily relies on global shape and the spatial arrangement of parts. Among the models tested, the Point Transformer exhibited a recognition style more akin to human behavior, primarily due to its hierarchical downsampling mechanism that enables the model to construct increasingly abstract shape representations. Ablation studies revealed that this hierarchical abstraction is crucial for integrating information across the entire shape, significantly enhancing the model's robustness in shape recognition. These insights point to a promising avenue for developing future AI models better equipped to recognize 3D objects in ways that mimic human perceptual abilities, offering potential improvements in various applications such as robotics and computer vision.