Oblique Imagery

Introduction - This week the lab task was to create 3D models out of processed imagery in Pix4D. Pix4D uses oblique imagery to create the models. This means that as a drone takes multiple pictures of one object, Pix4D uses the various angles to better represent the object in 3D.

The imagery was provided by Dr. Joseph Hupy at the University of Wisconsin - Eau Claire was from 3 separate flights of a DJI Phantom 3. One data set is of a small structure adjacent to a South Middle School (also in Eau Claire) track field (track shed), another is of a tractor located at the Litchfield Mine (imagery used in previous labs), and one of his old truck, the late "Guzzler," may she resell in piece.

Methods - Pix4D was the only GIS software used in this lab. The software was able to process the imagery, annotate the background, create point clouds and meshes, and create videos of the models.

First, all three flights underwent initial processing (Figures 5 and 6). Once initial processing was done, some of the images needed to be annotated, a process the tells Pix4D what is in the back or foreground of the imagery. Annotating (Figures 2, 3, and 4) various pictures from the data set trains Pix4D as to what is to be included and excluded out of the models. In this lab, the only parts of the images that weren't annotated were the track shed, tractor, and the truck, respectively (Figures 7, 8, and 9).

Four pictures were chosen from each data set to annotate, one from each side of the object so Pix4D would have annotated data from multiple angles. Annotating paints those areas of the images pink before applied. This is a tedious process that is quicker to do the further zoomed out the image is oriented. However, one can be more accurate zoomed in, and some areas required it, such as the truck wheel in Figure 4.

Once annotated, the imagery is then reprocessed with the point cloud and mesh (Figure 1), to include annotations. Once processed again, the imagery should create a clearer 3D model than what was produced in initial processing.

Figure 1: Pix4D processing point cloud and mesh.

Figure 2: Annotating a tight corner along the outside of a building

Figure 3: Annotated border (start) of object.

Figure 4: Zoomed in, correcting over annotated area over tire.

Figure 5: List of images available to conduct annotation

Figure 6: Angles of images in relation to object after initial processing

Figure 7: House with annotation

Figure 8: Tractor with annotation

Figure 9: "Guzzler" with annotation

Conclusion - There is a definite difference in the quality of the product. There are more areas that appear to be fuzzier than others without annotation. Below, (Figures 10 and 11) the difference can be seen. Figure 11 is much clearer than Figure 10 and has a much smoother boarder along the edges.

The same goes for the tractor animations, whereas the first video is without annotation and the second has annotation. The difference is very clear, in favor of the annotated imagery.

Had more pictures been trained, the expectation would be that the results would be even better. Though the process was tedious, the increased clarity in the product is worth the time. The annotating process was easy to do as well, and may be an ideal task for GIS interns.

Figure 10: House processed without annotation

Figure 11: House processed with annotation