A binary representation for real-valued, local feature descriptors

The usage of real-valued, local descriptors in computer vision applications is ofen constrained by their large memory requirements and long matching time. Typical approaches to the reduction of their vectors map the descriptor space to the Hamming space in which the obtained binary strings can be efficiently stored and compared. In contrary to such techniques, the approach proposed in this paper does not require a data-driven binarisation process, but can be seen as an extension of the floating-point descriptor computation pipeline with a step that allows turning it into a binary descriptor. In this step, binary tests are performed on values determined for pixel blocks from the described image patch. In the paper, the proposed approach is described and applied to two popular real-valued descriptors, SIFT and SURF. The paper also contains a comparison of the approach with state-of-the-art binarisation techniques and popular binary descriptors. The results demonstrate that the proposed representation for real-valued descriptors outperforms other methods on four demanding benchmark image datasets.

---

Opracowanie ze środków MNiSW w ramach umowy 812/P-DUN/2016 na działalność upowszechniającą naukę (zadania 2017).