A comparison of qualitative and metric spatial relation models for scene understanding

摘要: Object recognition systems can be unreliable when run in isolation depending on only image based features, but their performance improved taking scene context into account. In this paper, we present techniques to model and infer object labels real scenes a variety of spatial relations - geometric features which capture how objects co-occur compare efficacy the augmenting perception classification real-world table-top scenes. We utilise long-term dataset office table-tops for qualitatively comparing performances these techniques. On dataset, show that more intricate techniques, have superior do not generalise well small training data. also using coarser information perform crudely sufficiently standalone scenarios conclude expanding insights gained through comparisons comment few fundamental topics with respect autonomous robots.