![](bg2.png)
Some of the holes may also represent a real physical
hole, which is part of the structure of the object
being reconstructed. Unfortunately, just looking at
the point cloud, it is almost impossible to
differentiate between these two types of holes (real
and virtual). Particularly, when there is a real
physical hole in the structure of the target object
having uniformly distributed inner-color or inner-
details not appearing in the 2D views, this kind of
regions will appear as regions missing depths in the
estimated model. These are real holes that should be
left as they are during the filling process.
The ambiguity in differentiating between real
and virtual holes is a significant problem, which is
usually faced when running a hole-filling algorithm.
Thus, analysing each of the detected holes and
classifying them to either real or virtual hole is a
very important component in hole-filling process.
Finally, it is now clear that the problem of hole-
filling in point clouds requires two tasks. First,
identifying the holes, and then classifying them. By
classification, the necessity of filling the identified
hole can be determined. Unfortunately, both tasks
are nontrivial. But in this paper, we propose a simple
approach which will contribute to the accurate
detection and classification of holes in point clouds,
and consequently support systems for surface
reconstruction and hole filling.
The rest of this paper is organized as follows:
section 2 presents a brief background and lists the
related work. Section 3 discusses the proposed
approach. Experimental results and discussion can
be found in section 4. Finally, section 5 concludes
with a conclusion, perspectives, and future work.
2 BACKGROUND AND RELATED
WORK
Structure from Motion (SfM) techniques have
attracted researchers since the work of (Tomasi and
Kanade, 1992). SfM algorithm involves finding
correspondences between different input images for
the object being reconstructed, and then estimates a
3D model and a set of camera parameters.
Therefore, reconstructing regions having low
textures is challenging for most of SfM-based
approaches.
As mentioned before, the need for accurate 3D
models makes hole-filling a very important problem.
For a successful hole-filling, accurate detection and
classification of holes in point clouds are needed.
According to our knowledge, and after an intensive
review of the literature, not much work has been
published about detecting holes in point clouds.
Nevertheless, some methods employed either special
equipment or triangular meshes, sometimes
associated with some input entered manually by
users for hole-identification. For example, in (Noble
et al., 1998), the internal geometry feature of 3D
objects is measured by employing an X-ray
inspection method; thereby they were able to
position the drilled holes on the object's surface. In
(Kong et al., 2010), a hole-boundary identification
algorithm for 3D closed triangle mesh is presented.
In this method, the user has to interactively select
the region of interest by mouse dragging. From our
point of view, this method has one more drawback
besides the need for manual inputs. Dependence on
meshes instead of point clouds will not guarantee the
detection of all holes, because some meshing
algorithms may fill the regions of missing depth.
Consequently, this prevents distinguishing real from
virtual holes.
The authors of (Wang et al., 2012) proposed a
method which aims to find solid holes inside 3D
models. This method is also based on triangular
mesh models. By grouping interconnected coplanar
triangles, they extract the contour of the model using
the boundaries of the adjacent planes. Then, based
on the extracted contour, they form several disjoint
clusters of model vertices. Finally, by analysing the
relationship between the clusters and planes, holes
are identified. But, this method only finds solid
holes inside models, and does not detect regions of
missing depths.
The main goal of (Wang et al., 2007) is filling
holes in locally smooth surfaces. But, as a pre-
processing step, holes are found based on the
triangular mesh of the input point cloud. In this
method, holes are identified automatically by
tracking boundary edges. If an edge belongs only to
a single triangle, then it is a boundary edge,
otherwise it is a shared edge, which shares more
than one triangle. But, employing this strategy in
finding holes will detect all holes including the real
holes those need not be filled. Therefore, in this
work, user input is required as an assistance. So,
again, manual user inputs are needed in this work.
An automatic hole-detection approach has been
presented in (Bendels et al., 2006). Properties of
point sets have been investigated to derive several
criteria which are then combined into an integrated
boundary probability for each point. This method
seems to be robust, but we have noticed that it has
some drawbacks. First, in their combination of
probability criteria, they used some weights, which
VISAPP 2017 - International Conference on Computer Vision Theory and Applications
322