Multiple-Kernel Local-Patch Descriptor

Arun Mukundan,Giorgos Tolias,Ondrej Chum

DOI: https://doi.org/10.48550/arXiv.1707.07825

2017-07-25

Abstract:We propose a multiple-kernel local-patch descriptor based on efficient match kernels of patch gradients. It combines two parametrizations of gradient position and direction, each parametrization provides robustness to a different type of patch miss-registration: polar parametrization for noise in the patch dominant orientation detection, Cartesian for imprecise location of the feature point. Even though handcrafted, the proposed method consistently outperforms the state-of-the-art methods on two local patch benchmarks.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to represent and match local features more effectively in computer vision tasks. Specifically, the author proposes a multiple - kernel local - patch descriptor based on efficient match kernels (EMK), aiming to combine two parameterization methods to improve the robustness against different types of patch misregistration. These two parameterization methods are:

1. **Polar Parametrization**: mainly used to deal with the noise problem in patch dominant - direction detection.
2. **Cartesian Parametrization**: mainly used to deal with the imprecise feature - point positions.

By combining these two parameterization methods, the method proposed in the paper performs excellently in two local - patch benchmark tests, especially in the newly introduced HPatches dataset and the Phototourism benchmark test, significantly outperforming the existing methods.

### Specific Problems
1. **Improving the Robustness of the Descriptor**: Traditional hand - crafted descriptors have deficiencies in dealing with patch misregistration, especially for the noise in dominant - direction detection and the imprecision of feature - point positions. The paper improves the robustness of the descriptor against these misregistrations by combining the polar and Cartesian coordinate parameterization methods.
2. **Surpassing Existing Methods**: Although deep - learning methods have achieved good performance on some tasks, their generalization ability on different tasks and datasets is limited. The method proposed in the paper performs excellently on multiple tasks and datasets, and in particular, further improves the performance through supervised whitening.
3. **Reducing Computational Complexity**: The method proposed in the paper not only has superior performance but also has low computational complexity and can be quickly trained and applied on modern computers. In contrast, deep - learning methods require a large amount of computational resources and time.

### Main Contributions
- **Multiple - Kernel Local - Patch Descriptor**: Combines the advantages of polar and Cartesian coordinate parameterization, improving the robustness and performance of the descriptor.
- **Supervised Whitening**: Performs post - processing of the descriptor through supervised learning, further improving the performance of the descriptor.
- **Experimental Verification**: Conducts extensive experimental verification in multiple benchmark tests, proving the effectiveness and superiority of the proposed method.

Through these improvements, the paper provides a new and efficient local - feature - descriptor method, which is suitable for a variety of computer - vision tasks.