Uniqueness Metric for Comparison of Dense Image Descriptors
Mohsine Taarji, Arie Nakhmani · SoutheastCon 2022 · 2022
Image features and their descriptors are broadly used in many computer vision applications ranging from image retrieval and camera calibration to object recognition and image registration. Desired qualities of image features and descriptors frequently include low sensitivity to noise and invariance to illumination, scale changes, and rotation, while preserving uniqueness. Having non-unique or very similar descriptors at different regions of the same image might affect negatively feature matching algorithms and limit their usefulness. Presently, the choice of the best feature detector-descriptor pair for a particular application is based on comparisons using binary evaluation metrics such as accuracy and precision. While these metrics in many cases predict well the overall performance, they cannot provide granular information about the local behavior. Moreover, if the descriptor is dense, i.e., every pixel produces a feature vector, then using binary global metrics is meaningless. We propose a uniqueness metric, scaled between [-1,1], for descriptors. It allows identifying how unique they are in a predefined local neighborhood. The proposed metric also allows an easier exploration of noise effects on image descriptors. Our results show the applicability of the proposed metric to any dense descriptor and demonstrate how uniquely each pixel can be identified within the neighborhood. The proposed metric also could serve for the development of new feature detectors that optimize uniqueness. Based on that metric, we provide a comparison of state-of-the-art descriptors with regard to noise robustness.