Piecewise Planar Underwater Mosaicing Fabio Bellavia, Marco Fanfani, Fabio Pazzaglia and Carlo Colombo CVG, University of Florence Florence, 50139, Italy email: name.surname@unifi.it Riccardo Costanzi, Niccol` o Monni, Alessandro Ridolfi and Benedetto Allotta MDM Lab, University of Florence Florence, 50139, Italy email: name.surname@unifi.it Abstract—A commonly ignored problem in planar mosaics, yet often present in practice, is the selection of a reference homography reprojection frame where to attach the successive image frames of the mosaic. A bad choice for the reference frame can lead to severe distortions in the mosaic and can degenerate in incorrect configurations after some sequential frame concate- nations. This problem is accentuated in uncontrolled underwater acquisition setups as those provided by AUVs or ROVs due to both the noisy trajectory of the acquisition vehicle – with roll and pitch shakes – and to the non-flat nature of the seabed which tends to break the planarity assumption implicit in the mosaic construction. These scenarios can also introduce other undesired effects, such as light variations between successive frames, scattering and attenuation, vignetting, flickering and noise. This paper proposes a novel mosaicing pipeline, also including a strategy to select the best reference homography in planar mosaics from video sequences which minimizes the distortions induced on each image by the mosaic homography itself. Moreover, a new non-linear color correction scheme is incorporated to handle strong color and luminosity variations among the mosaic frames. Experimental evaluation of the pro- posed method on real, challenging underwater video sequences shows the validity of the approach, providing clear and visually appealing mosaics. I. I NTRODUCTION Video mosaicing [1] is a quite popular tool in underwater vision exploration and autonomous navigation [2], [3] since it gives a global, immediate, detailed and handy overview of the seabed, without requiring more expert knowledge in the 3D manipulation of the scene. However, despite the recent progress in the field [4], [5], obtaining good mosaics is still a challenging and not fully solved task. This is mostly due to the theoretical requisites of the mosaics, assuming input data with sufficiently distance from the scene or images acquired by camera rotations only [6]. These assumptions are not always met in practice, resulting in image misalignments and ghosting artefacts for which different blending techniques have been proposed [1], [7], [8], some of which tailored for the environments [9]. Furthermore, a commonly ignored problem, yet often present in practice, is the selection of a reference homography reprojection frame to which to attach the various mosaic images. The most common and trivial choice is to use the first image frame or, often supported by geo-referential camera positions, a user predefined one. A bad choice for the reference frame can lead to severe distortions in the mosaic and can degenerate in incorrect configurations after some sequential frame concatenations (see Fig. 1). This problem is accen- tuated in an uncontrolled underwater acquisition setup. For example videos acquired from AUVs or ROVs due to the noisy trajectory of the acquisition vehicle – with roll and pitch shakes – and, in most cases, the non-flat real nature of the seabed, tend to break the mosaic requirements. Moreover, these scenarios can introduce other undesired effects, such as light variations between successive frames, scattering and attenuation, vignetting, flickering and noise. In order to over- come these issues, recent underwater mosaic techniques also integrate 3D Structure-from-Motion and acoustic data [10], [11] to improve the results. However, these methodologies require more controlled environments or more hardware than a single camera. Fig. 1. A distorted mosaic due to a wrong reference homography selection (left), opposing to the proposed method (right). In both cases no post- processing color correction or blending have been applied to highlight the image frames. This paper presents a novel mosaic pipeline which includes a strategy to select the best reference homography in planar mosaics from video sequences. This reference homography globally minimizes the distortions induced on each image frame by the mosaic homography itself. Additionally, a new non-linear color correction scheme is presented which robustly handles strong color and luminosity variations among the mosaic frames.