Mobile robots need a compact and expressive representation of space to operate effectivelyin real world environments. We present here:
- The category-independent lower layers of a hierarchical representation of space that is based on laser range data, which has a potential to scale well.
- A low-level image descriptor, by which the performance of our representation is demonstrated in the context of a room categorization problem.
A large, freely available, Domestic Rooms Dataset has been collected, which is intended for room categorization experiments based on data obtained with a laser range finder.
#Spatial Hierarchy of Parts
A compositional hierarchical representation of space is learned from observations (Figure 1) and is comprised of several layers of parts (Figure 2). We present here the lower layers that are category-independent and are based on unsupervised learning of statistically significant observations, in terms of frequency of occurrence of various shapes in the environment.
The proposed approach was compared to the state-of-the-art CNN trained specifically for place recognition. The baseline experiments demonstrate that both methods achieve comparable performance on original scene images. Further experiments revealed that our method outperforms the holistic CNN by being robust to image degradation, such as occlusions, modifications of image scaling, and aspect changes.
#Histogram of Compositions:
Each laser scan is transformed into an image, from which elements of the hierarchy (parts) are inferred. These are used to form several histograms, where each of them corresponds to some particular region of the image. Their concatenation represents the Histogram of Compositions descriptor (Figure 3), which is used as the input for the Support Vector Machine classifier.
#Two Scenarios for Room Categorization
Two scenarios for room categorization have been considered. In the first scenario, called exploratory room categorization, the categorization is performed based on a set of laser scans obtained in each room, while in the second scenario, called single-shot room categorization, the categorization is performed based only on a single scan.
Using only the lower layers of the hierarchy, we obtain state-of-the-art categorization results on demanding datasets (Figure 4).
