Page 7 - i1052-5173-31-9
P. 7
georeferenced coordinates of the four mark- removed after alignment, improving the qual- Model) built in 2016 using an image survey
ers. This file was imported into Metashape, ity of the 3D scene reconstruction. These captured from the same outcrop with a
which allows the direct georeferencing of the images were identified through manual selec- dSLR camera (Fig. 4C). In this regard, the
model. The whole procedure, from the export tion of points associated with unrealistic or same fault was mapped in 2016 (Corradetti
or unregistered data from Metashape, blurry geometries within the sparse cloud. et al., 2021), using 640 images (4272 × 2848
through the rotations, scaling, and referenc- Often those were frames characterized by pixels) taken with a Canon EOS 450D reflex
ing in OpenPlot and the final re-import in extreme overlap. mounted on a tripod to suppress motion
Metashape takes just a few minutes and can Both point clouds are characterized by blur. The reconstructed area for the Reflex
be followed step-by-step in the supplemen- zones on their boundaries, in which the 3D Model was ~2.67 m , and the point cloud
2
tary video provided (see Supplementary scene reconstruction relies on oblique images included ~2.7 × 10 points. These three point
8
Material ). A good practice consists of check- (Fig. 4B). These zones are asymmetrical, clouds were uploaded in CloudCompare
1
ing the results and re-exporting the cameras’ due to the aforementioned obliquity between (Girardeau-Montaut, 2015), where they
extrinsic data of the registered model to pos- the fault-perpendicular direction and the were first manually aligned using ~15 con-
sibly repeat the procedure if residual rota- average photo view direction. Accordingly, trol points for each matched point cloud,
tions occur (i.e., if ρ is not perfectly lying on we cropped the point clouds to exclude and then they were compared using the
a horizontal plane), which may relate to the these zones and areas where the 3D recon- cloud-to-cloud distance tool. The resulting
proximity of the markers used for the trans- struction relied upon less than nine images distance among the three clouds was gener-
form and on their positional accuracy. (Fig. 4B). ally below 4 mm (Fig. 4D), which decreases
The cropped point cloud for the Photo down to <2 mm for the Photo Model versus
RESULTS Model is composed of ~2.5 × 10 points, Reflex Model.
7
For the Photo Model, all of the 200 uploaded whereas the cropped Video Model consists The georeferenced Photo and Video mod-
photos were successfully aligned and used to of ~7.8 × 10 points (Fig. 4C). The accuracy els were then compared to evaluate differ-
7
produce a point cloud made of ~6 × 10 points of these 3D surface reconstructions was ences in scaling and rotation (translation
7
(Fig. 4A). For the Video Model, we uploaded tested by generating difference maps from was not investigated here). To achieve this,
735 video frames, but only 528 of them were the two smartphone-generated models, and we uploaded the two scaled and rotated
successfully aligned and used to produce a between each smartphone-generated model models, using the compass holder as the ori-
dense cloud of ~11.6 × 10 points (Fig. 4A). and a high-resolution ground-truth model gin of the reference frame. We aligned the
7
Some of the excluded images were manually (from here on referred to as the Reflex two clouds using 15 control points, and the
Figure 4. The Photo and Video models dense point cloud (A). (B) Images positions with respect to the models and number of images overlapping areas. (C)
Cropped Photo and Video models. The Reflex Model from Corradetti et al. (2021). (D) Cloud to cloud distance between each pair of point clouds computed
in CloudCompare.
1 Supplemental Material. Video of the registration procedure in Metashape and OpenPlot. Metashape reports. Go to https://doi.org/10.1130/GSAT.S.14751042 to access the
supplemental material; contact editing@geosociety.org with any questions.
www.geosociety.org/gsatoday 7