Estimating Cluster Stability in Adaptive Resonance Theory for XR Image Understanding

Clustering techniques play a crucial role in scene understanding in extended reality (XR) applications such as user perspective analysis and visual localization. Fuzzy Adaptive Resonance Theory (Fuzzy ART) is a fast and convenient clustering method; however, its performance is significantly affected by the preset vigilance parameter. Although existing improvement methods incorporate mechanisms for adjusting the vigilance parameter, their complex algorithmic frameworks or processes often hinder the transparency and usability of the algorithms. To address this issue, we retain the simplicity and ease of use of Fuzzy ART and propose Stability-Based Vigilance Adjustment Fuzzy ART (SV-ART). Specifically, SV-ART integrates the Stability Assessment (SA), Vigilance Adjustment (VA), and Extinct Cluster Deletion (ED) modules into the iterative process of Fuzzy ART, uncovering implicit knowledge throughout the iterations to personalize the fine-tuning of vigilance values for each cluster, thereby optimizing the clustering structure through continuous iterations. We conducted experimental evaluations on 12 datasets and found that SV-ART improves both the optimal and overall clustering performance. This demonstrates that SV-ART improves its sensitivity to the vigilance parameter while retaining the robustness of Fuzzy ART. This makes it particularly suitable for non-expert users in scenarios where prior knowledge is scarce.