Iterative Reconstruction of Large Scenes Using Heterogeneous Feature Tracking

With image capturing technology growing ubiquitous in consumer products and scientific studies, there is a corresponding growth in the applications that utilize scene structure for deriving information. This trend has also been reflected in the plethora of recent studies on reconstruction using robust structure from motion, bundle adjustment, and related techniques. Most of these studies, however, have concentrated on unstructured collections of images. In this paper, we propose a feature tracking and reconstruction framework for structured image collections using heterogenous features. This is motivated by the observation that images contain a small number of features that are fast/easy to track and a large number of features that are difficult/slow to track. By tracking these separately, we show that we can not only improve the tracking speed, but also improve the tracking accuracy by using a camera geometry based descriptor. We demonstrate this on a new challenging dataset which contains images of Arctic sea ice. The reconstruction pipeline constructed using the proposed method provides near real time reconstruction of the scene, enabling the user to parse vast amounts of data rapidly. Quantitative comparisons with baseline SFM techniques show that reconstruction accuracy does not suffer.