Far from knowing it all when it comes to photogrammetry, Will had to cut his teeth working hard to make some sense of the thing before producing the imagery, models and videos the Soggy Wombats have been sharing in social media and presentations. An early presentation to a small audience of CDAA members (Cave Divers Association of Australia) was done to share some of the knowledge gathered through shear brute force working with Agisoft Metashape. Not a software issue, more a user inability to operate it, but the results speak for themselves. The intent of this post is to share not only the slides presented back in September 2024, but also some of the information exchanged in the day. Under the true Wombat spirit and mission of sharing knowlege as much as we can to make cave diving safer and more enjoyable to everyone, this is for the ones curious about how these things are created and the work involved.

Photogrammetry can be defined as “the science and technology of obtaining reliable information about physical objects and the environment through the process of recording, measuring and interpreting photographic images and patterns of electromagnetic radiant imagery and other phenomena.” The Soggy Wombats didn’t make this one up. This is how photogrammetry is defined in Wikipedia.

Although more recently the scuba diving community has been exposed to a lot of photogrammetry applications, it is not a ground breaking ‘spanking new’ technology. Photogrammetry research goes back to 1867 – roughly as old as ‘modern’ photography (1822). While the invention of the method is attributed to Aimé Laussedat, the term “photogrammetry” was coined by the Prussian architect Albrecht Meydenbauer, which appeared in his 1867 article “Die Photometrographie.”

Applications of the technology such as 3D laser scans, point cloud, dense cloud and BIM are stock standard in the construction industry and among surveyors. Everyone have done at least one walkthrough on a GloogleMaps or a real estate virtual visit. The larger range of applications means there is a vast amount of software available as well as documentation for reference. The specific underwater application we as scuba divers are developing is the big news here. Access to cheaper and higher quality digital cameras, accessible software, and more powerful home computers have all created the right mix to boost the ‘domestic’ use of the technology.

The whole process is based on still photos. The software will ‘map’ all the photos in a matrix of 40,000 points for example, with overlapping images being the key for the process to work. Any photogrammetry software will only validate a point if it is detected in more than a single photo/image. The more images with the same point in them, the higher the accuracy of that point’s position. More info/photos is always better, but only to a point. The software expects alternative angles of the same point, so if you give it 400 photos from the same position, it is quite a waste of time and data.

There is a lot of information on how to scan a physical 3D object, like a cube or a running shoe or whatever you might want to make a digital 3D model of… however, caves are kind of the opposite of an ‘normal’ 3D object. While if you want to scan a cube, you would make it the central focus of your photos and move the camera around it to gather different angles and perspectives, in a cave you are at the center of it and the cave is around you. This makes the process of gathering data a bit more complex than if you would trying to scan a simple 3D object. Don’t panic, as I mentioned before there is a good amount of free literature on how to use photogrammetry to scan caves. Just not underwater…

As a Project Manager (yes that is what pay the bills) I know planning is always how start any project. And you can’t plan anything without a little understanding of what is you final goal. So we start backwards. What do we want to achieve? To improve safety and awareness of the cave environment while we are underwater. Why do we want this? To improve planning and training before any dive even starts. How can we achieve this? By developing a 3D model and making it available. Well, the loop is then closed.

Some of the Wombats annoyed me for quite a while to get into it, but before I had time to think, plan and be happy with it, nothing happened.

At first, I’ve learned the ropes at home. To simulate the environment I would find in underwater caves I hade blinds down, phone flashlight on and two GoPros Hero 3 taking photos every second. You don’t want to see my place, but the results were very interesting. I’ve noticed immediatelly that the more lights you can have the better. Issue at home were the flat white walls… the software hates it! So the intensity of the light you are throwing at the surfaces plays a big part on the final result, just like regular photography. If you have a bunch of super exposed photos, your scan will be bad. But it’s hard to control all these different surfaces and exposure levels underwater…

For my first test run underwater, I’ve picked some photos from a photo shoot in Pines Cave. Hear me out, instead of going for a dive to grab photogrammetry data, I’ve done the opposite: Do I have enough photos from a ‘section’ of any cave that I can use? And the answer was yes. The entry/exit of the TV Room in Pines is a point where I did a few photo shoots with the Wombats and other diving buddies. I had around 8 shots with someone getting in or out of the TV Room – results can be seen in slide #4 top centre and right photos. Photos were taken with a Canon 5D MKIII full frame camera with double strobe firing in sync.

It felt like I had a test case to validate. As long as I was able to have a good camera with good light, I should be able to to get decent results.

The Internet is a great researching source, and finding other people around the world who were developing underwater photogrammetry projects wasn’t that hard. I’ve contact my now friend Marcin Stepmpniewicz as he was one of the people posting some very cool things about the Maria Concordia Project in Poland. His results were amazing and I felt confident to exchange some of my first findings with him from the TV Room in Pines Cave. One of his recommendations was to use GoPros, which at the time I thought would not be as good as the images collected with the DSLR. So it was time to check both set-ups in a real environment.

Allendale Cave in the Mount Gambier region is very easy to access and provides a great training ground for inexperienced cave divers in Australia. I thought that this was the perfect place to test the things I’ve learned and the hypothesis I was making about data gathering and results in that cave. I did two dives with a similar circuit throught the cave each time. The first dive with a rig with four GoPros Hero3 and two BigBlue 7200 lumens video lights. The second dive I used my Canon 5D MKIII paired with two synced strobes.

The results were night and day. The quality of the point cloud and textured model created by the DLSR (Canon 5D MKIII) was ridiculously good when compared to the GoPros. Texture, level of detail, colour… every single aspect of the final result was far superior. However, I could not cover the entirety of the cave with the DLSR as the system was too big to squeezy it past a very simple and not too narrow restriction in Allendale Cave. We don’t dive through restrictions all the time whe diving in a cave, but when you do find them the last thing you want is to have to negotiate your extremely large camera set-up through it while trying to take photos from different angles… final veredict: GoPros are the go.

Believing I’d tamed the process, it was time to keep moving forward. Iddlebiddy Cave was selected as the next site to get a full scan. Larger in both tunnel diameter and length, with multiple jumps and some very silty areas, Iddlebiddy would present a greater challenge to my still limited experience with photogrammetry. The rig setting that worked perfecly to Allendale produced an absolutely terrible result for Iddlebiddy. It felt like I had not enough light and not the right cameras. An update to four new GoPros Hero 11 and an extra BigBlue 7200 lumens unit fixed part of the issue though. The last piece of the puzzle was the positioning of the cameras among themsleves. Scanning a tunnel-like cave is different to scanning a room-like cave… and the worst part is that you can only tell all of this after you drive 5 hours from Mount Gambier back to Melbourne and spend several days in front of your computer at home.

Knowledge exchange was key to solving the challenges at Iddlebiddy. A few conversations with Marcin helped me understand what modifications would give me the greatest benefit, and how different camera position changes would impact the results.

The final challenge to tame was scanning Pines Cave. “The Pines” has a mix of very large rooms (like Allendale) and lots of tunnel sections (like Iddlebiddy), so I felt I was ready to combine the lessons and face it. As it usually happens, I was absolutely wrong. The biggest challenge in Pines Cave wasn’t the dark underwater sections, but the massive main cavern zone with a very large surface pool. This pool received direct sunlight and the transition between ‘no-lights’ to ‘sunlight’ brought a whole new set of challenges.

I’d started playing with photogrammetry using an old iMac late 2013 3.4Ghz Quad Core i5 / 24GB RAM / NVIDIA GTX 775 2GB and 1TB 7200rpm HDD – certainly not the fastest machine available in 2023. I needed to upgrade my kit anyway, so decided to go with MacStudio at the time, and I’ve updated again since then. Today I run all my photogrammetry workflow out of a MacBook Pro M3 Max 16CPU – 40 GPU 64GB RAM, 2TB SSD.

You don’t need a Mac to run photogrammetry though, and I think I might be one of the few people that does. You don’t need a brand new PC either, your old computer might be completely capable of running Agisoft, it might just take longer to spit the results out (but please still check the software minimum requirements).

Whatever you decide to use, disk storage and RAM will be your main restrictions. RAM has reduced its price substantially, but it is still probably the most expensive thing you can buy for any computer. And for photogrammetry, you want as much as you can pay for. The number of photos you generate in a cave scan will depend on how much detail you want in your model/texture. Don’t break the bank, but don’t expect to be able to have a functional workflow with 4GB RAM. Fast storage within your computer (as close as possible to the CPU and GPU) will be key to make your models be generated faster, but you may then need a separate solution to store the final model. The number of images you will have to store might surprise you initially, so make sure you have a reliable external storage solution so you don’t have to keep everything in your computer’s high-speed internal disks.

The number of cameras used in data collection is also a function of a few things. How many dives do you want to do in order to get the imagery required for the level of detail you’d like in the final model? Are the dives easily accessible and can you go back anytime? If it’s tough and access is limited, then I would take as many cameras as possible on the one tough in the dive. Better to have too many images than to have not enough images and no easy way to collect more…

The number of lights you will need also depends on the environment you are facing. I found that three 7200 lumens are enough for the conditions I encounter in the Mount Gambier caves. Water visibility, the darkness of walls in the underwater cave, how far the lights placed away from the cameras… so many variables and it’s too hard to give you a simple answer that works for everyone everywhere.

However Agisoft Metashape seems to be the global standard for everyone working with underwater photogrammetry. There is a multitude of software available and if you are familiar with one, I would encourage you not to learn a new one. I knew nothing about it and tried a couple different ones, with Metashape feeling the most user-friendly and the most commonly referenced by other underwater users. I’m not sponsored by Agisoft and bought my own license, opting for the Pro license as it allows you to work with geo-referenced photos and use markers. The geo-referencing is very important when I’m producing maps and overlaying the model onto satellite imagery from GoogleMaps. The result is so much better as I can give the user some real world references. Markers are also extremly useful in my workflow when I’m joining two date-sets, especially as underwater and above-water photo sets usually are hard to ‘tie’ together, but markers can solve this problem in seconds instead of requiring weeks of processing.

Above is an example of my typical workflow. I’m getting better at not repeating dives as I gain more experience, but there is always the chance that I’ll blunder a dive and ‘waste’ time gathering photos that will not be useful for a photogrammetry model. Due to the size of Pines (the cave shown above in “Slide #6 – My Workflow”) and the level of detail I wanted to get, I planned multiple dives with clear goals to achieve within to specific areas of the cave. I did not want to have to scan the 42m depth area while trying to collect shallower data as well. Another important point is to try to limit the scans to areas that you can join together with previous data, or to ensure overlap with data that you’ll collect on a later scan. This data overlap helps considerably, as when you get something out of the computer it feels like you are ‘building’ something by connecting the pieces together.

Drone imagery is also fundamental for me. Every scan has the ultimate goal of being represented in a 2D map of the cave in focus. Without surface geo-referenced information, I can’t create the maps that I do. Even if we can’t fly the drone, GoPros can collect GPS data as well to pinpoint the final model in the real-world. The drone imagery has also helped many times to set the correct size/scale of the model by enabling the measurement of known objects on the surface such as cars, handrails, and staircases.

Most recently we have been using Insta360 images to create some of the models we have been sharing. It is a fantastic tool for the job. Once again, position of the light sources is very important. And although it will help get a shape and direction of the cave in a fraction of the time the four GoPros set of data would, the level of detail of the textured model is far inferior when compared to the GoPros results. Depending on what you are looking for, both solutions are correct and viable, but they will give you completely different outcomes. Do your planning, and use the best tool for the job.

From the initial Allendale results (where all the red dots are low accuracy points) to the Pines result (where the red dots are far less frequent) there was quite a lot of trial and error, with vital learnings drawn from every project and cave.

The above info was shared several years and we have learned much more since. The Soggy Wombats have been invited to numerous dive conferences to speak about underwater cave photogrammetry, and every time we are on stage we try to share the latest project, model, scan or video with our audience. Frequent updates to our website, Instagram and Facebook pages are the other way we choose to share our latest efforts with a wider audience. Keep your ears to the ground and like/follow us on social media if you’d like to know what we have been up to most recently.

I hope this helps those curious about underwater photogrammetry and encourages them to try it themselves. Feel free to reach out if you need any guidance or just to exchange some thoughts and/or information – sharing is caring.