Beyond SIFT for image classification

In classifying images, scenes or objects, the most popular approach is based on the features extraction-coding-pooling framework allowing to generate discriminative and robust image representations from densely extracted local patches, mainly some SIFT/HOG ones. The majority of the latest research is focused on how to improve successfully these coding and pooling parts. In this work, we show that substantial improvements can be also obtained by coding information closer to the pixel values level in the same way that deep-learning architectures do. We introduce a two layer, stacked, coder-pooler architecture where the first layer is specifically dedicated to extract, from our so-called Differential Vectors (DV) patches, some efficient, local low-level features more discriminative and efficient that their classic handcrafted counterpart. This first layer can advantageously replace any classic dense SIFT/HOG patches extraction stage. We demonstrate the effectiveness of our approach on three datasets: UIUC-Sports, Scene 15 and Caltech 101. We achieve excellent performances with simple linear classification while using basic coding and pooling schemes for both layers, i.e. Sparse Coding (SC) and Max-Pooling (MP) respectively.