Multiple View Stereo with quadtree-guided priors

A survey on conventional and learning‐based methods for multi‐view stereo

“…Unfortunately, reliable depth estimates cannot be efficiently calculated in particularly low discriminative areas, even with such a robust propagation. Many efforts have been made in this direction, either by using more efficient propagation and sampling patterns (Xu & Tao, 2019), coarse‐to‐fine approaches (Wei et al, 2014; Xu et al, 2020; Xu & Tao, 2019), semantic priors (Stathopoulou et al, 2021) or direct planarity constraints (Romanoni & Matteucci, 2019; Stathopoulou et al, 2023; Xu & Tao, 2020b), achieving improved results.…”

“…Semantic priors have also been integrated in the PatchMatch MVS pipeline to support depth estimation in challenging areas and enable class‐selective 3D reconstruction (Stathopoulou et al, 2021). To lessen the requirement for a priori generated labels, Stathopoulou et al (2023) adopted a quadtree image decomposition scheme to detect local planes and propagate more reliable depth estimates where matching ambiguities occur.…”

Adaptive region aggregation for multi‐view stereo matching using deformable convolutional networks

“…In this context, many researchers have exploited the a priori structures inferred from the images or from additional inputs. For instance, planar structures from images (Stathopoulou et al, 2023; Xu & Tao, 2020), an additional digital elevation model in the object space (Ling et al, 2016), a filter window based on surface orientation constraints (Huang & Qin, 2020), similar segments established by a minimum spanning tree (Wang & Xu, 2023; Yang, 2012), a self‐adaptive triangle constraint or line segments in the images (Qin et al, 2019; Ye & Wu, 2018; Zhu et al, 2010) can facilitate the determination of the supporting domain for image matching. The window can also be constructed using non‐symmetrical shifts (Bobick & Intille, 1999) that are meticulously selected by avoiding the cross‐sectional regions of the object contours (Wu et al, 2012), the assumptions of the a priori structures are not universal, and the determination of shiftable windows is non‐trivial.…”

“…The photo‐consistency (Remondino et al, 2014; Stathopoulou et al, 2023), which indicates the pixel‐wise similarity between a set of images, must be accurately determined for image matching. In general, dense pixel‐wise descriptors are constructed, and a proper distance metric is then used to estimate the similarities.…”

“…Ideally, the estimation of the photo‐consistency must be confined within a certain region, for example, within the boundaries of the target object. Low‐level geometrical features, such as line contours (Bobick & Intille, 1999; Wu et al, 2012), planar segments (Stathopoulou et al, 2023) or superpixels (Ji et al, 2021), are typically used to adaptively select the supporting domain. However, low‐level features are vulnerable to noise and may not correspond to the object boundaries.…”