UPDATE 28 Oct 2012: Fixed missing files in code distribution. The first version had files missing in two submodules.
In this work, we explore two ways to customize a nearest-neighbor database specifically for each query.
Our first method learns per-descriptor weights to minimize classification error. In addition to adjusting the relative weight of different types of descriptor, this also has the effect of throwing away distractor and outlier examples formed by imperfect superpixelation and descriptor generation.
Our second method prunes the training set used for each query based on image context, conditioning on common classes to mine the database for extra rare class examples. This re-balances the class frequencies, away from the highly-skewed distribution found in real-world scenes.
The classification framework is inspired by the work of work of Tighe and Lazebnik, who use a simple kNN scheme with multiple descriptor types to classify super-pixels.