Quantify NetVLAD association robustness in visually repetitive environments

Quantify the robustness of NetVLAD-based keyframe encoding and submap association in visually repetitive environments, where incorrect submap associations may occur.

Background

CGS-SLAM uses NetVLAD descriptors to compare keyframe encodings across agents and identify spatially corresponding keyframes for submap alignment. The supplementary experiments report no incorrect submap associations on the evaluated datasets, but those datasets do not establish how the association mechanism behaves in environments containing repeated or highly similar visual structures. The unresolved issue is therefore to evaluate its robustness under visually repetitive conditions, which are particularly relevant to reliable multi-agent reconstruction.