| shell grit 
nsity pro- 
urposes a 
| Or Views 
' environ- 
erspective 
TOm vari- 
ng profile 
below the 
 sparse T- 
1 amongst 
  
  
d T-MVS 
itly below 
(b)). This 
'ormation. 
in a future 
produced 
)isson sur- 
is 2.3 mil- 
louds (re- 
fter clean- 
nd ~6000 
>ss the ef- 
e it to the 
te a dense 
e overlaid 
x 4 d 
n A > 
(a) The blue UAV-MVS points amongst the T-MVS points from 
above. 
(b) The blue UAV-MVS points beneath the T-MVS points from below. 
Figure 7: A 6 cm wide strip (profile A) of the UAV-MVS point 
cloud viewed with the T-MVS point cloud. 
The wider 6 cm profile strips have been created as surfaces to as- 
sess the difference between UAV-MVS Poisson and T-MVS Pois- 
son. Figure 3 shows the TIN surface compared to the Poisson 
surface for T-MVS and UAV-MVS datasets respectively. 
In these views the natural coloured point clouds have been offset 
in the Z dimension by -10 cm to allow visualisation of the shape 
of the surface compared to the cloud that it was derived from. The 
surface covers a vegetated section and in the UAV-MVS dataset 
the denser section previously mentioned can be seen when com- 
paring this view (Figure 8(a)) to the same view of the shrub (Fig- 
ure 8(b)). The T-MVS cloud is sparse here and as a result the 
Poisson surface seems to have exaggerated the shrub height over 
the sparse section, probably due to the orientation of the normals 
varying greatly for those few points, which happens in vegeta- 
tion. The triangulated mesh is much more jagged than the Poisson 
surface in both views and the UAV-MVS Poisson is particularly 
smooth (Figure 8(a)). The shrub in reality does have a reasonably 
smooth shape, in in this instance the UAV-MVS Poisson appears 
most accurate. To examine this further a section of Profile B that 
passes through the pebbly beach is visualised. In Figure 9(b) the 
Poisson surface is again smoother and the drop in terrain at this 
point point is well represented (see Figure 6(b)). In Figure 9(a) 
the same seems evident. In Figure 9(c) the Poisson surfaces for 
UAV-MVS and T-MVS are shown on the T-MVS point cloud (Z- 
10 cm). In this view the UAV-MVS surface is again ~1 cm below 
the T-MVS surface, but the shape of the terrain is basically the 
same, where as when two raw MVS based TINs are compared in 
the same view the outliers in the UAV-MVS data seem to cause 
the surface to vary suddenly causing spikes or peaks in terrain 
that are not evident in the equivalent T-MVS TIN. 
These visualisations provide insight into the quality of terrain and 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
479 
   
  
(a) The UAV-MVS point cloud below Poisson (blue) and TIN 
(light blue) strips. 
  
A 
(b) The T-MVS point cloud below Poisson (brown) and TIN 
(pink) strips. 
Figure 8: 6 cm wide strips of Poisson and TIN surfaces viewed 
over a vegetated section of the points clouds from which they 
were derived (Z-10 cm), each natural coloured dot has a 14mm 
diameter. 
surface extraction possible using MVS techniques. The use of 
Poisson surface reconstruction has potential advantages over tra- 
ditional triangulated mesh creation. Poisson surfaces seem gen- 
erally smoother and smooth surface representations are often bet- 
ter when undertaking decimation, hydrological analysis, DEM 
derivative extraction and vegetation and ground filtering. The 
apparent outliers in the point cloud may not impact on the out- 
puts from these analyses and, provided the point cloud density is 
carefully monitored and taken into account when mapping sur- 
face quality, the result may be more realistic for most surface 
types. Some vegetated areas have complex geometry (such as 
complex overlapping branches or tussock grasses) and areas with 
little or no texture are going to be poorly represented and this 
may impact on the Poisson reconstruction. The creation of a TIN 
is still a viable option, particularly when the point cloud can be 
maintained without decimation. When products with a smaller 
memory footprint are required there seems to be a strong case for 
using Poisson surface reconstruction to create a fairly smooth yet 
detailed representation of the terrain from which lower resolution 
surfaces can be extracted. The TIN surfaces appear more jagged 
and these spikes in the terrain can cause erroneous height values 
in a derived output. 
4 CONCLUSIONS AND FUTURE WORK 
This study presented a qualitative assessment of the accuracy of 
point clouds derived using multi-view stereo techniques (MVS).