In part 2 of my series on photogrammetry, I will discuss the process of creating the three-dimensional model of the bowl from the National Museum (for more information on the process of taking the images of the bowl—which I mentioned in my last post—see part 1 of Explorations in Photogrammetry). The process itself involves the use of two pieces of software: Adobe Photoshop and Agisoft Photoscan Pro.
Photoshop to the Rescue
The first step was to unify the images. This process was done using Photoshop and a few of its “auto” features. I began by setting up a batch job (learn how to create batch jobs in Photoshop here). This job applied the following auto corrections in order:
- Auto contrast – to adjust the brightness settings of each photograph and create a uniform brightness/contrast
- Auto colour – to correct any oddities in colour
- Auto tone – to smooth out an residual white / black in the image and apply a universal look
Each file was then re-saved in a separate location. It’s best practice to always keep a copy of the original image, unedited.
Building the Model in Photoscan Pro
The next step was to import the photos into Photoscan Pro. Photoscan Pro is software that allows a number of images to be “stitched” together using a point cloud. This point cloud is then used to construct a 3D model of the object.
The first step involved assessing the quality of the images. Photoscan Pro has a built-in quality assessment tool. After running this tool against all the images imported into the programme, each photo is then given a score (from 0 to 1) that shows how high the quality of the image is in order to align it and produce a 3D model. While learning how to use the software in class, we were taught a general rule of thumb is to only use images whose quality is .6 or higher.
After importing the images and assessing their quality, I received the following results:
|Number of Images||Quality||% of Total|
|7||.40 – .49||5%|
|45||.50 – .59||29%|
|101||.60 and above||66%|
I was a little dismayed that such a large number of my images (over 30%) were under the .60 threshold. I decided to run two different models in order to compare the results.
I began by excluding all images under .60 quality. I applied a mask to each image so as to instruct Photoscan to ignore everything other than object (this involved basically “cropping” out the background and having Photoscan ignore anything but the bowl object—a very time intensive process but well worth the results). I then used Photoscan to align the photos, build a point cloud, a dense point cloud, and then a mesh. This created a somewhat reasonable 3D model of the bowl; however there were a number of errors in its rendering and the model itself looked incomplete.
In order to correct this issue, I attempted to rerun the model by including the 45 images that were marked with an image quality of .50 – .59. When including these images, along with the original 101 images of .60 quality or better, I received a much stronger model. This model lacked the errors in its rendering that were present in the first model and looked more complete. Upon close inspection, however, one can see the very bottom of the bowl looks as though it was “cut out”. This is not a feature of the bowl itself, but rather a failure to capture images at a deep enough angle so as to capture the entire inside of the bowl.
As I began to closely inspect the second version of the bowl, I also noticed there was quite a bit of “noise” in the model (areas of the bowl that looked pixelated or out of place). I attempted to reapply the masks on the objects by cropping the images further (thus excluding the very edges of the object in each photo). However, this did not turn out well. My bowl ended up rather flat-looking, as though someone had collapsed the bowl upon itself. I think took the original point cloud from the second model and began manually cleaning up the cloud by looking for areas of white (which were the background and thus not part of the bowl). After manually removing these points, I rebuilt the mesh and texture and developed my final model.
Overall, the bowl turned out better than I initially expected, especially given this was my first attempt at such an endeavour. Given the opportunity to repeat the process under the same settings, I would make the following corrections:
- Adjust the aperture and ISO. For most of the images, I was using an F-stop of 8 and an ISO of 640. I think I would adjust this to lower the ISO (thus reducing visual noise in the image) and raise the F-stop on the camera. The exact values I would use are difficult to say, but I would most likely try to lower the ISO to somewhere between 100 and 200 and raise the F-stop up to at least 22. This would give me a crisper image, especially in regards to depth of field, as well as the added benefit of removing a small amount of noise that exists in the image.
- Adjust the shutter speed. I left the shutter speed on the automatic setting for this exercise, mainly because I didn’t think I would need to adjust it. Given another opportunity, I might attempt to manually adjust this setting to see how it affects the quality of light in the image (many of images seemed a little dark, despite all of the artificial lighting present).
- Adjust the angle. I would also take another round of photographs at a much steeper angle so as to completely capture the interior of the bowl.
- Take more time. As I was the first student to begin taking pictures, I self-imposed pressure to complete my pictures quickly, so as to ensure my fellow classmates had ample time with the camera as well. If I repeated this process again, I would be less conscious of time constraints and take more time in between photos to evaluate the quality of the image and adjust my settings as needed.
In an effort to show off the difference between the two different models, I’ve included both below, as well as a version of the bowl that was sent to the 3D Printer (more on 3D printing in a future blog post). The first is the “bad” model that only included the images with a quality of .60 or higher. The second model includes images with a quality of .50 or higher. Also note the absence of part of the bottom of the bowl where I failed to take images at a deep enough angle to cover the entire bottom of the bowl. The model used by the 3D printer does not have the hole in the bottom of the bowl as I used 3ds Max to fill in the mesh.
Special thanks to the National Museum of Ireland and their Archaeology department for allowing us to photograph some of their objects. The pictures taken of the bowl featured in this post and my previous post were courtesy of National Museum of Ireland – Archaeology. All rights reserved.
Coming Up Next…
In part 3 of my series on photogrammetry, I will discuss my choice for the second part of my assignment. I will detail why I chose the particular object in question and what challenges were presented to me as part of this aspect of the assignment. I will also detail the steps I took to overcome those challenges.