Reducing Drift in Mosaicing Slit-Lamp Retinal Images

Active Annotation of Informative Overlapping Frames in Video Mosaicking Applications

“…After dense alignment, a threshold on the loss function is then used as heuristic to accept or discard the performed registration. Following this seminal work, other approaches used the current reconstruction as an indicator of the topology of the sequence (Can et al, 2002;Kang et al, 2000;Loewke et al, 2011;Marzotto et al, 2004;Prokopetc and Bartoli, 2016). Since the cost of bundle adjustment updates between iterations can be problematic, Gracias et al (2004) proposed an affine bundle adjustment formulation allowing online updates of the bundle adjustment results directly on the mosaic.…”

“…However, in practice, other transformation spaces with less degrees of freedom are often used instead. In increasing order of complexity, examples of such alternative transformation models include translations (Mahé et al, 2013), rigid transformations (Richa et al, 2014), similarity transformations (Elibol et al, 2013;Garcia-Fidalgo et al, 2016;Molina and Zhu, 2014) and affine transformations (Choe et al, 2006;Gracias et al, 2004;Peter et al, 2018;Prokopetc and Bartoli, 2016). Occasionally, transformations with more than 8 degrees of freedom can be used to account for applicationspecific non-planarity effects, such as in retinal imaging (Can et al, 2002;Yang and Stewart, 2004).…”

“…In this work, similar to Choe et al (2006); Gracias et al (2004); Peter et al (2018); Prokopetc and Bartoli (2016), we use affine transformations. This choice is primarily motivated by the fact that affine bundle adjustment can be formulated as a linear least-square problem and that, as a result, new pairwise correspondences can be efficiently added via a recursive linear least-squares formulation (Gracias et al, 2004).…”

“…A well-known example of mosaicking task is the panoramic image stitching found in standard digital cameras (Brown and Lowe, 2007). Mosaicking algorithms are particularly needed in a variety of biomedical applications such as retinal imaging (Can et al, 2002;Choe and Cohen, 2005;Prokopetc and Bartoli, 2016;Richa et al, 2014), fetoscopy (Reeff, 2011;Tella-Amo et al, 2018), fibered endoscopy (Atasoy et al, 2008), confocal endomicroscopy (Loewke et al, 2011;Mahé et al, 2013;Vercauteren et al, 2006) and 3D ultrasound imaging (Ni et al, 2008;Wachinger et al, 2007). Applications of video mosaicking are also typically found in remote sensing applications, for example to reconstruct the seabed from underwater sequences (Gracias et al, 2004;Elibol et al, 2014) or to obtain a high-resolution map of a city from a set of aerial images (Kekec et al, 2014;Molina and Zhu, 2014;Xia et al, 2017).…”

“…Indeed, finding annotable pairs requires to identify and remember revisited locations in the sequence, a task which is far from trivial in visually challenging environments. For these reasons, mosaicking algorithms are notoriously difficult to evaluate quantitatively (Atasoy et al, 2008;Prokopetc and Bartoli, 2016). Apart from synthetic and phantom experiments for which a gold standard is more easily available, mosaicking algorithms in "real-world" scenarios are often only evaluated by qualitative visual assesssment of the rendered mosaic (Atasoy et al, 2008;Mahé et al, 2015), or by using as a surrogate gold standard the result of a more computationally intensive but more reliable algorithm such as bundle adjustment (Kekec et al, 2014).…”