Cascade of Box (CABOX) Filters for Optimal Scale Space Approximation

Local image features, such as blobs and corners, have proven to be very useful for several computer vision applications. However, for enabling applications such as visual search and augmented reality with near-realtime latency, blob detection can be quite computationally expensive due to numerous convolution operations. In this paper, we present a sparse convex formulation to determine a minimal set of box filters for fast yet robust approximation to the Gaussian kernels used for blob detection. We call our feature detector as CABOX (CAscade of BOX) detector. Although box approximations to a filter have been studied in the literature, previous approaches suffer from one or more of the following problems: 1) ad hoc box filter design, 2) non-elegant trade-off between filter reconstruction quality and speed and, 3) limited experimental evaluation considering very small datasets. This paper, on the other hand, contributes: 1) an elegant optimization approach to determine an optimal sparse set of box filters, and 2) a comprehensive experimental evaluation including a large scale image matching experiment with about 16 K matching and 170 K non-matching image pairs. Our experimental results show a substantial overlap (89%) between the features detected with our proposed method and the popular Difference-of-Gaussian (DoG) approach. And yet CABOX is 44% faster. Moreover, the large scale experiment shows that CABOX closely reproduces DoG's performance in an end-to-end feature detection and matching pipeline.