Local Similarities, Global Coding: An Algorithm for Feature Coding and its Applications
This work addresses a fundamental problem in image processing by enhancing coding methods for tasks like classification or recognition, though it appears incremental as it builds on existing algorithms.
The authors challenged the locality assumption in feature coding by proposing a method that incorporates global similarities among anchor points using a random walker, showing that this approach improves various state-of-the-art coding algorithms across multiple datasets.
Data coding as a building block of several image processing algorithms has been received great attention recently. Indeed, the importance of the locality assumption in coding approaches is studied in numerous works and several methods are proposed based on this concept. We probe this assumption and claim that taking the similarity between a data point and a more global set of anchor points does not necessarily weaken the coding method as long as the underlying structure of the anchor points are taken into account. Based on this fact, we propose to capture this underlying structure by assuming a random walker over the anchor points. We show that our method is a fast approximate learning algorithm based on the diffusion map kernel. The experiments on various datasets show that making different state-of-the-art coding algorithms aware of this structure boosts them in different learning tasks.