**6. Conclusions and Future Works**

This paper proposes two different methods to compact topological maps. With this aim, three datasets from indoor environments were used. These datasets were composed by either panoramic images or omnidirectional images that were transformed to panoramic. During the experiments, with the objective of compacting the information, the number of instances was reduced to a value in the interval from 10–100. That means a reduction of instances up to between 1.1% and 11.5% of the original number. The proposed methods were (1) spectral clustering and (2) self-organizing maps. Moreover, three global appearance descriptors were used since they presented a good solution for environments whose data dimensionality was high. The work shows that it is possible to reduce the visual information drastically from the original model. Among these combinations of method-descriptor, spectral clustering along with the *gist* descriptor was proven to be the best choice to compact the model.

Once the original model is compacted, the resultant map can be used to solve the localization task. Hence, an evaluation is carried out with the aim of measuring the goodness of the localization task through the use of compact maps and global appearance descriptors. In this case, three descriptors and two indoor environments are evaluated. Furthermore, a mixture between indoor environments is created with the aim of evaluating whether it is possible, first, to detect the right environment and, second, estimate the position of the instance. From this study, HOG is the description method whose localization results were the best. Additionally, *gist* presented the most successful results in order to select the correct environment of a test instance from a combined dataset. Finally, the use of clustering methods to tackle the compression step has proven to be more efficient than carrying out a downsampling of the images directly from the database.

The team is now working on how the localization task through compact maps is affected by illumination changes. Additionally, other compacting methods will be studied in order to achieve the Simultaneous Localization And Mapping task (SLAM).

**Author Contributions:** Conceptualization, L.P. and O.R.; Methodology, L.P. and W.M.; Software, S.C.; Validation, L.P., S.C. and W.M.; Formal Analysis, O.R. and L.P.; Investigation, S.C. and O.R.; Resources, L.P. and W.M.; Data Curation, S.C. and W.M.; Writing—Original Draft Preparation, S.C.; Writing—Review & Editing, L.P. and O.R.; Visualization, S.C. and O.R.; Supervision, L.P.; Project Administration, O.R.; Funding Acquisition, L.P., O.R. and S.C.

**Funding:** This research was funded by the Generalitat Valenciana through Grant ACIF/2017/146 and by the Spanish government through the project DPI2016-78361-R (AEI/FEDER, UE): "Creación de mapas mediante métodos de apariencia visual para la navegación de robots".

**Conflicts of Interest:** The authors declare no conflict of interest.
