Deadlines and Detail
Having masses of data is a double edged sword. Too little and your images won’t align or your models lack definition. Too much and processing time jumps exponentially. Its always a balancing act between a desire for detailed models and the time available.
When The Thistlegorm Project team decided to launch the website on the 76th anniversary of the sinking a hard deadline formed. Decisions regarding time to process vs model had to be taken and concessions were made – Mike Postons at 3Deep Media needed time to work his magic on the models and delivering a really detailed model the day before go-live was not an option.
Thankfully Agisoft Photoscan Pro has many settings that can be tweaked and applied during the model build. Photo alignment, dense cloud processing, meshing and texture processing have values to adjust. Setting the processing values are always balance of available time vs desired quality.
Identification and Counting
With more time we can start to take a look at refining areas of the 3D model. The holds and their cargo are such a space.
Alex Mustard has done a lot of work to name the vehicles but not all the cargo has been identified. Objects in the holds remain unidentified.
As the models built it made sense to build small sections with finer detail and have a closer look. This technique helped confirm the IDs of vehicles, and count the motorbikes with confidence.
Run it Again
With more time and no deadline I could re-run areas of interest again.
Here’s a screenshot from the original model of hold 2 deck 1, starboard side. There are rows of BSA motorcycles stacked in Morris-Commercial trucks:
Example of the 3D model of hold 2 deck on, starboard side. The dense cloud, model and texture parameters were all set on the low side. The result is things look fine from a distance but blocky and pixilated when viewed close up.
Building the model of deck 1 took 5165 images and several days of processing using the lower quality values and settings. The result was a model good enough for Mike to take and use.
But I knew there was more detail to be had.
Restricting processing to just 772 images of the starboard side of hold 2 took 24 hours to process but the higher quality settings applied at every stage and the result was a big improvement over the first pass model:
Refined example of the 3D model of hold 2 deck on, starboard side. The dense cloud, model and texture parameters were all set on the high side. Note the improvement in detail in the motorbike mudguard stays, tyres etc.
The actual model is online for viewing. At 4.5m faces its not the quickest to load, but its worth waiting for. Some of the wheels have individual spokes visible amongst the rust and marine detritus:
Have Data. Will Process.
Producing accurate models still begins with a well exposed and in focus 2D photograph. With the images on the card the photogrammetry work really starts and like any journey its a lot easier if the end result is known. Storing photogrammetry data does mean a steady stream of hard drive space is needed but once you have the data refinement is possible.
What Next?
There is a lot of the Thistlegorm just begging to be reprocessed. The bow section of hold 1 deck 2 needs a second look as its cargo is unidentified. But what area would you like to see refined and reworked with more detail? Feel free to use the Contact Us page to get in touch and share your thoughts.