PeerAppear: A Location-Aware Framework for Extensible Image Annotation and Peer-to-Peer Discovery

This paper addresses the problem of building and maintaining image repositories which form the basis for world-scale visual models. The availability of these models enable capabilities such as visual localization, persistent surveillance, and structure from motion. We approach this problem through the creation of PeerAppear, a location-aware framework for extensible image annotation and peer-to-peer dissemination. Due to the dynamic nature of the world, solutions to this problem generally require significant effort to be spent on mapping. PeerAppear enables a decentralized solution through the implementation of a peer-to-peer middleware framework which automates the indexing and sharing of visual information extracted from images in a user's collection. The PeerAppear network achieves scale through a Bittorrent style overlay network, which indexes the locations of user's image collections using a hierarchical geographic segmentation scheme in a distributed hash table. By enabling a distributed and collaborative solution, network participants are able to provide for frequent remapping and a search capability which leverages efficient and effective visual representations. The framework was implemented in Python and evaluated using Raspberry Pi computers for data collection and a workstation computer to simulate network participants. This evaluation showed that the framework was able to provide an effective image search capability and shows promise for large-scale operations.