Descriptor evaluation and feature regression for multimodal image analysis

Abstract

In this paper, we present feature descriptor evaluation and feature regression for multimodal image analysis. First, we compare the performances of several popular interest point detectors and feature descriptors from multimodal images with focus on visual and infrared images. The performances of detectors are evaluated mainly by the score of repeatability and accuracy and the descriptors are assessed by using the rate of precision and recall. Secondly, we analyze the relationship between the corresponding descriptors computed from multimodal images. The descriptors are regressed by means of linear regression as well as Gaussian process. Then the features on infrared images are predicted by mapping the descriptors from visual images to the infrared modality through the regression results. Predictions are assessed in two ways: the statistics of absolute error between true values and actual values, and the precision score of matching the predicted descriptors to the original infrared descriptors. We believe that this evaluating information will be useful when selecting an appropriate detector and descriptor for multimodal image analysis. Also the experimental results show that regression methods achieve a well-assessed relationship between corresponding descriptors from multiple modalities.