Table 2. Overview of mosaicing approaches for fibred endoscopic imaging.

Topic	References	Methodology	Comments
Image based, real-time	Bedard et al. (2012) and Vercauteren et al. (2008)	Local rigid alignment through normalised cross correlation matching.	Simple, local, rigid registration based on image similarity maximisation.
	Loewke et al. (2008)	Local rigid alignment through feature based optical flow refined via gradient descent on normalised cross correlation.	Provide valuable real-time feedback during data acquisition for effective mosaic generation.
			Certain assumptions and model simplifications required for achieving real time performance.
Image based, post-procedural	Vercauteren et al. (2005,2006)	Hierarchical framework of frame-to-reference transformations (on the original, sparsely sampled data) to derive a globally consistent rigid alignment, while compensating for motion distortions, elastic deformations. Global alignment seen as an estimation problem on a Lie group.	More complex models dealing with a range of local and global, rigid and elastic image transformations. Post-procedural approaches with real-time capacity compromised due to the underlying complex registration models.
	Loewke et al. (2011,2007a)	Compensating for global (rigid) as well as local (elastic) transformations (including motion distortion. Fixed correspondence between images were replaced with a Gaussian Potential representing the amount of certainty in the registration. Global and local deformation potentials were maximised in an integrated optimisation problem
	Hu et al. (2010)	Elastic registration of consecutive frames based on optical flow of robust image features (RANSAC strategy on Lucas-Kanade tracker. A Maximum a Posterior (MAP) estimation based image blending generated super-resolved images.
Image based, dynamic imaging	Mahé et al. (2015)	Dynamic mosaic obtained by solving a 3D Markov Random Field. Two-stage approach, of static mosaicing followed by stitching of the associated video segments.	Generating mosaics that maintain temporal information in the form of infinite loops.
External input based	Loewke et al. (2007b)	Initial rigid alignment using feedback from a robotic arm determining the five degree-of-freedom position/orientation of the fibre tip.	Actuators/sensors provide feedback on the scanning path improving the efficiency and/or the robustness of mosaicing. Hardware additions are limiting their suitability for endoscopic applications.
	Vyas et al. (2015)	Initial rigid alignment using feedback from a six degree-of-freedom electromagnetic sensor positioned at the tip of the fibre-bundle.
	Mahé et al. (2013)	Weak a-priori knowledge of the trajectory (spiral scan) used to derive spatio-temporal associations within the frame sequence. A hidden Markov model notation and a Viterbi algorithm was recovered the optimal frame associations, feeding a modification of the mosaicing algorithm by Vercauteren et al. (2006) to estimate the optimal transform.