Self help: Seeking out perplexing images for ever improving navigation

Self help: Seeking out perplexing images for ever improving navigation,10.1109/ICRA.2011.5980404,Rohan Paul,Paul Newman

Self help: Seeking out perplexing images for ever improving navigation  
BibTex | RIS | RefWorks Download
This paper is a demonstration of how a robot can, through introspection and then targeted data retrieval, improve its own performance. It is a step in the direction of lifelong learning and adaptation and is motivated by the desire to build robots that have plastic competencies which are not baked in. They should react to and benefit from use. We consider a particular instantiation of this problem in the context of place recognition. Based on a topic based probabilistic model of images, we use a measure of perplexity to evaluate how well a working set of background images explain the robot's online view of the world. Offline, the robot then searches an external resource to seek out additional background images that bolster its ability to localise in its environment when used next. In this way the robot adapts and improves performance through use. I. INTRODUCTION This paper is about having a robot actively seek data to improve its understanding of the world. The big picture motivation of the work is to enable robot longevity and in this paper we consider a specific instantiation of this problem - that of asymptotically improving scene recognition with a camera. We shall make use of the FAB-MAP algorithm (Cummins et al. (4), (5)) which probabilistically associates a current view of the world (image) taken by a robot with a previously visited or a new place. FAB-MAP requires priming with a set of images (which we shall refer to as a sample set) which in concert, statistically represents the appearance of the robot's workspace. For operation in urban settings, for example, one equips it with a sample set containing random images of cities and towns. There is an obvious shortcoming here - the robot is constrained to work in settings in which its sample set has sufficient explanatory power. If moved into surroundings quite different from those represented by its sampling set performance drops - nothing is as expected and everything is astounding. In this paper, we show how by producing a generative model of the underlying topics present in observed images we can actively grow a customised sample set by incorporating well chosen examples from an external corpus which is more representative of the workspace the vehicle is experiencing. In this way, we replace the inflexibility of a static a-priori sample set with a plastic, dynamic one and we show this affords an improvement in performance over time. One could think of this as a robot actively seeking to widen its experience, better understand its surroundings and becoming less perplexed with time. Our problem setup is as follows. A mobile robot must maintain a compact on-board sample set summarizing the visual appearance of its environment. The robot explores the environment collecting image data and identifies the most perplexing images based on its current sample set. It then searches the least explained images in a large repository of image data (or past datasets from the robot) finding images with similar thematic content. Next, the robot retrieves examples (based on their likelihood in the environment) and assimilates them into the sample set, thereby improving its representation and performance. The rest of the paper details this framework and presents the following components: Use of Latent Dirichlet Allocation (LDA) topic model to extract a low-dimensional thematic representation for images incorporating word co-occurrence statistics in an unsupervised manner, Section III. Identifying most novel images seen by the robot given its current representation using a perplexity-based mea- sure, Section IV. Finding images similar in thematic content from an external repository applying language-model based in- formation retrieval approach, Section III. Application to FAB-MAP, presenting an algorithm for constructing a representative sampling set, Section V. Experimentation on real datasets collected by a mobile robot, Section VI.
Cumulative Annual
View Publication
The following links allow you to view full publications. These links are maintained by other sources not affiliated with Microsoft Academic Search.