Reproduce “Human-Inspired Topological Representations for Visual Object Recognition in Unseen Environments” on OpenResearch

About this paper

Ekta U. Samani, Ashis G. Banerjee

Researchers at the University of Washington, Seattle, developed THOR2, a framework that integrates human-inspired topological representations of 3D shape and color for visual object recognition. Trained solely on synthetic data, THOR2 achieved 75.07% recognition accuracy on the OCID dataset and 62.58% on the UW-IS Occluded dataset, outperforming a state-of-the-art RGB-D ViT model and its shape-only predecessor, THOR.

Key insights

• Topological data analysis, when combined with perceptually uniform color spaces, can create robust, low-dimensional, and view-invariant representations that generalize effectively from synthetic to diverse real-world environments.
• Leveraging human-inspired principles like 'object unity' allows a recognition system trained on unoccluded objects to successfully identify objects under partial occlusion in unseen real-world scenes, significantly reducing real-world data dependency.
• The fusion of spatially-aware topological shape and color features, integrated through a color similarity network, offers superior robustness to sensor noise, varying lighting conditions, and partial occlusions compared to single-modality or purely data-driven deep learning approaches.

Summary by alphaXiv.