Improving multi-view image classification using higher order information and triangulation embedding

In this paper we take a look at extensions of the Bag of Words model developed within the last few years. Namely the aggregation of vector residuals known as VLAD encodings and Fisher kernels and assess their performance for the classification task using multiple views. We also take a look at the triangular embedding strategy for classification in the compression domain. Our work focuses on using the binary descriptor known as ORB. We are also able to show that triangular embedding is extremely fast and can provide the best performance without direct spatial encoding on the images themselves and we also demonstrate a novel approach to improve the triangulation accuracy that is less prone to overfitting than the traditional approach. We are able to show that higher order information provides improved performance in the multiple view setting. We finally take a look at the use of multiple view classification for fast image classification with a variable number of views and that our approach using a modified triangular embedding can overcome information loss and be used real-time.