As we mentioned in the previous post, we are investing a lot of our time on analysing point clouds. An important parameter influencing the success rate and the accuracy of the 3D reconstruction is the amount of texture details on the object surface. To better understand the influence of this parameter on the structure from motion process, we conduct a lot of experimental work using a series real world case studies. Two of these case studies are presented below.
The first example is the deanery, built in 1773, in the municipality of Zomergem, Belgium. The point cloud was built with 48 images and contains 75,000 points. It clearly shows the shape of the building. Although the images were focussed on the main building, the majority of the points were not detected in the main building, but in the features surrounding it, in particular in the trees in the background and in the small wall and the hedge in front of the building. Also clearly visible is the white wall of the smaller building (to the left of the main building), while the white wall of the main building is ‘pointless’.
The main reason for the differences in the density of detected points are differences in the amount of features in the surfaces. The small brick wall, the hedge and the trees are feature-rich surfaces. A lot of points are detected in these. The wall of the small building is a white painted brick wall. Although it is an equally white surface, the presence of the bricks and the very small differences in height between the individual stones still allows the detection of points. The wall of the main building is a white plastered wall. This surface is clearly unsuitable for image-based 3D modelling.
The second example is a head (of a very kind volunteer). Interesting is the clear difference in point density between the face and the hair. The skin of the face proved to be very suitable for image-based 3D modelling, while the hair, because it is a shinny surface, proved to be unsuitable.
Case studies like these are very useful to determine and to understand the weaknesses of image-based 3D modelling, and to study how we can deal with these weaknesses when doing ‘the real work’.