Introspection in Learned Semantic Scene Graph Localisation
This work addresses interpretability in semantic scene graph localisation for robotics or vision systems, but it is incremental as it focuses on introspection and validation of existing methods.
The paper investigates how semantics affect localisation performance and robustness in a self-supervised, contrastive framework, finding that the model learns noise-robust, semantically salient relations for explainable registration under variations.
This work investigates how semantics influence localisation performance and robustness in a learned self-supervised, contrastive semantic localisation framework. After training a localisation network on both original and perturbed maps, we conduct a thorough post-hoc introspection analysis to probe whether the model filters environmental noise and prioritises distinctive landmarks over routine clutter. We validate various interpretability methods and present a comparative reliability analysis. Integrated gradients and Attention Weights consistently emerge as the most reliable probes of learned behaviour. A semantic class ablation further reveals an implicit weighting in which frequent objects are often down-weighted. Overall, the results indicate that the model learns noise-robust, semantically salient relations about place definition, thereby enabling explainable registration under challenging visual and structural variations.