Giovanni Cordara

Dual-exposure image registration for HDR processing

In this paper we present a method for registering a pair of differently exposed Low Dynamic Range (LDR) images for the purpose of rendering a High Dynamic Range (HDR) image. In general, the images are captured from a moving camera and/or contain moving objects. Therefore, proper registration is required to enable HDR rendering. However, even for equally exposed images, registration is an ill posed problem where errors are expected for a wide range of image pairs. The problem only becomes more challenging for a pair of differently exposed images. We propose an adaptive registration error detection and correction method to address this issue. By combining Optical Flow with the proposed correction method, we achieve state-of-the art results as shown in numerous experiments. The proposed method is simple and has low-complexity, hence allowing for an easy and efficient implementation.

Editorial: Special issue on visual search and augmented reality

Modern era mobile phones and tablets have evolved into powerful image and video processing devices, equipped with high-resolution cameras, color displays, and hardwareaccelerated graphics. They are also equipped with location sensors, and connected to broadband wireless networks, allowing fast transmission of information. This enabled a new class of applications utilizing phone’s built-in cameras to initiate search queries about objects in visual proximity to the user, thus providing fast and effective way to interact with the surrounding world. The enabling technologies are usually referred to as Visual Search and Augmented Reality: the former refers to the automatic identification of objects depicted in a picture, and the latter indicates the capability to convey to the user additional content related to the recognized elements. These applications have catalyzed interest of researches in computer vision community, and last few years have witnessed a remarkable progress, leading to first prototypes and commercial applications. Nevertheless, massive adoption of such technology is not on the way yet. Many technical factors are at the basis of this: computational efficiency and battery consumption, need for data compression, accuracy in retrieval and reliability in object tracking, lack of standard interfaces. These are some of the aspects that still demand for improvement of the current state of the art, and are object of study in the scientific community. This special issue aims at drawing a global picture about the research being carried out on visual search and augmented reality, and presenting some of the most compelling new developments addressing and improving the abovementioned aspects. The issue starts with two papers presenting methods for compressing the amount of information extracted from pictures to be sent over the networks. The nest two papers address the computational efficiency problem, presenting respectively a low complexity approach for local tracking of objects and an implementation exploiting GPUs in order to accelerate the searching phase. They are followed by two papers combining visual hints with other information available to improve accuracy and user experience. The special issue ends with the presentation of a dedicated ongoing standardization activity.

Automatic video to point cloud registration in a structure-from-motion framework

In Structure-from-Motion (SfM) applications, the capability of integrating new visual information into existing 3D models is an important need. In particular, video streams could bring significant advantages, since they provide dense and redundant information, even if normally only relative to a limited portion of the scene. In this work we propose a fast technique to reliably integrate local but dense information from videos into existing global but sparse 3D models. We show how to extract from the video data local 3D information that can be easily processed allowing incremental growing, refinement, and update of the existing 3D models. The proposed technique has been tested against two state-of-the-art SfM algorithms, showing significant improvements in terms of computational time and final point cloud density.

Statistical modelling for enhanced outlier detection

Matching of local features is an uncertain process which may provide wrong associations due to several reasons that include, among other factors, the uncertainty in locating the keypoint position. Since the statistics of the Log Distance Ratio (LDR) for pairs of incorrect matches are significantly different from those of correct matches, we propose a noniterative scheme for outlier detection that includes in the distance calculation the location uncertainty of the keypoint, specifically modeled by a covariance matrix: the LDR is then evaluated relying on Mahalanobis distance. By statistically modeling the wrong associations, inlier matches can thus be rapidly identified by solving an eigenvalue problem. The method is general enough to be applied both in 2D (i.e., texture) and 3D (i.e., texture + depth) scenarios. The effectiveness of the proposed method is assessed in the field of RGB-D SLAM, showing significant improvements with respect to state of the art methods.