Bayesian Surprise and Landmark Detection Ananth Ranganathan Frank Dellaert aranganathan@honda-ri.com dellaert@cc.gatech.edu Honda Research Institute, USA College of Computing, Georgia Inst. of Technology Cambridge, MA Atlanta, GA, USA Fig. 1. Top 20 SIFT feature patches by histogram count from the bag- of-words model for each location denoted by the measurement number for an experiment. Only every second measurement is shown. The measurements corresponding to landmarks (i.e. where the landmark detector fires) are shown in red (shaded overlay). It can be seen that these correspond to the start of sub- sequences of measurements that also differ qualitatively from the preceding measurements, for example measurements before 34 are much more cluttered than those following it. Abstract— Automatic detection of landmarks, usually special places in the environment such as gateways, for topological mapping has proven to be a difficult task. We present the use of Bayesian surprise, introduced in computer vision, for landmark detection. Further, we provide a novel hierarchical, graphical model for the appearance of a place and use this model to perform surprise-based landmark detection. Our scheme is agnostic to the sensor type, and we demonstrate this by implementing a simple laser model for computing surprise. We evaluate our landmark detector using appearance and laser measurements in the context of a topological mapping algorithm, thus demonstrating the practical applicability of the detector. I. I NTRODUCTION We introduce a novel landmark detection scheme, based on Bayesian surprise, for use in topological mapping. Our method detects landmarks as “special places” in the environment that can be added as nodes in the graph corresponding to the topological map. The notion of “surprise”, first proposed by Itti and Baldi [7], encodes the unexpectedness of a measurement and has been shown to be a good predictor of directed human attention [8]. Further, the computational framework for Bayesian surprise, which is based on a KL-divergence type measure, is quite simple and computationally efficient. For vision-based sensors, we introduce a new hierarchical graphical model, called the Multivariate Polya model, based on the bag-of-words paradigm. The model explains the common Fig. 2. Topological map, showing landmarks detected for the sequence of Figure 1 using Bayesian surprise (see also Figure 5). The smoothed trajectory is also shown. Nodes belonging to the same physical landmark have the same color. observation that if multiple images are taken at a particular place, their SIFT histograms are rarely exactly the same. Hence, SIFT histograms can be viewed as noisy measurements of the appearance of a place, and can be modeled accordingly. The Multivariate Polya model is known in the text-modeling community [1], but our usage of it in this context is novel. We derive the computation of Bayesian surprise using this model, which involves an approximation of the model to the exponential family of distributions. Our landmark detection technique generalizes to other sen- sors as well, and we demonstrate this by deriving a simple surprise model for laser range scans. Sensor-independence of our scheme is obtained through its computation in a Bayesian framework. Bayesian surprise supports the inclusion of measurements from multiple, distinct sensor sources, the only requirement being that a measurement model is defined for each of the sensors. We incorporate our landmark detection scheme into the topological mapping algorithm given by us in previous work [18] to produce a complete topological mapping system. Landmark detection is evaluated in the context of this mapping system using appearance and laser on a number of envi- ronments, including publicly available datasets that are well- known in the robotic mapping community. An analysis of the number of false negatives and false positives output by the technique is also presented. We present related work in the next section and define Bayesian surprise in Section III. This is followed by an explanation of how surprise is used in landmark detection in Section IV. In Section V, we briefly introduce the topological mapping algorithm used to evaluate the landmark detection