For some time, OpenSfM, the photogrammetry library maintained by Mapillary that underpins OpenDroneMap, has had support for 360 cameras. We are working on a project at the moment with some great engineering students from Case Western Reserve University on building a next generation 360 camera for photogrammetry, but while that project wraps up, I wanted to test what can be done with a commodity unit.
So, with hunker-down-in-place orders the du jour, I opted to do my initial tests in-between my house and the neighbors:
I know I probably don’t need my mask there, but I have nasty tree allergies, so I am taking advantage of the normalization of mask wearing to keep my lungs healthier than they are most Spring seasons.
First, the why-what?!
In the increasingly tenuously named OpenDroneMap project, we have seen some interesting alternatives to drones in use — general photogrammetry that I owe more blog posts on (in the meantime, you can sneak-peak them at https://smathermather.com — I have just been too busy to reblog them yet). From tiny pits and seeds of hard to identify plants to animal skulls, there are some interesting non-drone use cases for good photogrammetry.
Are drones enough?
Drone mapping is a really exciting and useful innovation: it allows for mapping large areas with low capital investment, an opportunity to leverage local talent, can often capture with a faster cadence, and higher resolution and has a small fossil fuel footprint as compared with using manned aircraft. But the detail available is not always the detail needed. Consider dense urban locales, especially in places that also are thickly vegetated, and drone mapping may not always be enough for capturing the bottom-of-the-urban-canyon elevations needed for certain detailed hydrological analyses.
360 → 3D?
With a 360 camera and enough walking, can we create a synoptic understanding of our world that augments what we are doing now with drones? Tests from my driveway are very promising.
One of our community members, Zach Ryall, has covered OpenDroneMap in a recent article for AOPA (Aircraft Owners and Pilots Association).
Maps and mosaics are among the most powerful products a drone camera can produce, but producers of polished, user-friendly software made for professionals aren’t giving freebies anymore. Open-source software offers hobbyists and soloists an affordable alternative.
Small correction on the attribution of credits from the article: OpenDroneMap was founded by Stephen Mather and not me (I’d like to get this record straight). If it wasn’t for Steve’s commitment and vision for the project, we wouldn’t be reading this post (or this article, or anything OpenDroneMap related for the matter).
Thanks to the tireless work of the folks behind the Zanzibar Mapping Initiative, I have been exploring the latest settings in OpenDroneMap for processing data over Stone Town. I managed to get some nice looking orthos from the dataset:
But, excitingly, I was able to extract some nice looking surface models from the dataset too. This required using the Brown-Conrady model that recently got added to OpenSfM:
This post is a small homage to the late his Majesty Sultan Qaboos. Given the strong affinity and shared history between Zanzibar and Oman, it seems fitting to post these.
Edit: the dreaded “Reconstructing all views” message has been replaced with a progress monitor! But, how to dig into the back end and explore the machine that does the work is always a helpful skill to have… .
We’ll learn how to check how far along a process is when it is calculating depthmaps and giving no feedback
Along the way, we’ll also learn how to look at the docker images running for WebODM and
login to said docker images in order to
inspect that status of processing data.
So, you threw a really big dataset at WebODM, and now you are waiting. It’s been hours, maybe days, and it’s stuck on the dreaded “Reconstructing all views”:
Did you make the depthmap resolution really high, because you wanted really detailed data? Did you make it too high? How long is this going to take?
I have had this dilemma too. Sometimes I just get disgusted with myself and my predilection for choosing ridiculously computationally expensive settings, kill the process, turn the settings down a bit, and restart. But this can waste hours or days of processing, which feels wrong.
We could do the alternative. We could poke under the hood of WebODM and see how it’s progressing. For this project that’s been running for 146+ hours, this is what I decided to do.
The depthmaps that are running are being done in MVE, which can give us really detailed information about progress, but unfortunately, it makes logs a real mess, so we have it logging nothing. Let’s see how we can get around this and check in our status.
Finding the docker instance and logging in:
First, we log into the machine where WebODM is running. We need a list of the existing docker machines, as we need access to the correct machine to look at how things are progressing.
The result should give us something like this:
CONTAINER ID IMAGE COMMAND CREATED STATUS PORTS NAMES
4b0659fe6761 opendronemap/webodm_webapp "/bin/bash -c 'chmod…" 38 hours ago Up 38 hours 0.0.0.0:443->8000/tcp, 0.0.0.0:80->8080/tcp webapp
0e26ebf918f2 opendronemap/webodm_webapp "/bin/bash -c '/webo…" 38 hours ago Up 38 hours worker
1954c5136d44 redis "docker-entrypoint.s…" 38 hours ago Up 38 hours 6379/tcp broker
bdc69502ca50 opendronemap/webodm_db "docker-entrypoint.s…" 38 hours ago Up 38 hours 0.0.0.0:32769->5432/tcp db
81f401a0e138 opendronemap/nodeodm "/usr/bin/nodejs /va…" 38 hours ago Up 38 hours 0.0.0.0:3000->3000/tcp webodm_node-odm_1
We want to access the webodm_node-odm_1 node, in most cases. To do this we use docker exec as follows:
Typically, if we only have one process running, there will only be one dataset in the /var/www/data directory
Checking our depthmap data from MVE:
For depthmaps nearly complete in MVE, there will be a file called depth-L1.mvei. We need to find out how many these are as compared with the number of images that we need depthmaps for. We’ll use a combination of the find command and wc (or word count):
find . -name depth-L1.mvei | wc -l
In my case, I have 2,485 images, or roughly 2/3s of my images processed. Looks like I am 6 days into a 9 day process before we get done with the MVE step.
I guess I will wait until Monday to check again… .
In a previous blog post, we explored how we can quite effectively derive terrain models using drones over deciduous, winter scenes. We ran into some limitations in the quality of the terrain model: the challenge was removing the unwanted features (things like tree trunks) while retaining wanted features (large rock features).
I concluded the post thusly:
For our use case, however, we can use the best parameters for this area, take a high touch approach, and create a really nice map of a special area in our parks for very low cost. High touch/low cost. I can’t think of a sweeter spot to reach.
Good parameters for better filtering
In the end, the trick was to extract as good of a depthmap as possible depthmap-resolution: 1280 in my case, set the point cloud filtering (Simple Morphological Filter or SMRF) smrf-window and smrf-threshold to 3 meters to only filter things like tree trunks, and set ignore-gsd: true to ensure we are keeping the highest quality data all the way through the toolchain.
WebODM introduced support for plant health algorithms about a month ago. It was no secret that we concurrently started work to support TIFFs file inputs and multispectral cameras, both features that have been highly requested.
Today we are excited to announce the release of ODM 0.9.8!
Up until now ODM was able to process only JPG files. With this release we added support for processing TIFF files, both 8bit and 16bit. TIFF is a popular format especially with multispectral and thermal cameras.
Multispectral Support (Experimental)
We have added the ability to process multispectral images (for example those captured with a Parrot Sequoia, MicaSense Altum or RedEdge). We have obtained some promising preliminary results. When provided with N camera bands, ODM will generate an N band orthophoto (in the proper bit resolution, up to 16).
We have not added support for spectral calibration targets, which we plan to add in the near future and we’re currently looking to add support for more cameras. The task of identifying different bands is different for each camera vendor and we’ll need to add support for more cameras with time. We hope the community will start to process some datasets and help us improve multispectral support (share your datasets?)
We recommend to pass the –texturing-skip-global-seam-leveling option when processing thermal/multispectral datasets. Global seam leveling will attempt to normalize the values distribution across orthophotos, which works well for making pretty RGB images, but will affect measured values in thermal/multispectral settings.
We rewrote from scratch the orthophoto cutline blending algorithm to merge orthophotos, a bottleneck that was causing processing to take longer than necessary in the last stage of the pipeline. The new algorithm is faster and much more memory efficient. We also sped up by a factor of 30x the time it takes to merge point clouds from submodels, as well as reducing memory usage drastically.
We brought the latest version of OpenSfM in this update, which delivers up to 1.6x faster image matching than before.
MVE has also been (finally) modified to report progress on the status of computations. You’ll finally know if the program is “stuck” or not.
Improved Brown-Conrady Camera Model
We made modifications to the camera model used to compute camera poses and points. Up until now the default in ODM has been to use a simplified perspective camera model. The community has been testing the usage of the brown-conrady model for a while, with great results. The original brown-conrady model however uses two parameters for specifying focal length, which are unfortunately not accurately supported by the texturing and dense reconstruction stages of the pipeline (a single focal length is used by those stages). We’ve approximated the brown-conrady computation by averaging the two focal lengths, but could we do better? Yes!
We modified the brown-conrady model to use a single focal length, bringing the model to full support for all stages of the pipeline and set it as the default camera model. We expect this will improve the quality of results for all outputs. Preliminary tests confirm this.
NodeODM now exposes a task list API endpoint. This is a feature that has been requested a lot and allows people to view the tasks running on a particular node. If you send a task to NodeODM via WebODM (or CloudODM or PyODM, or any other client), if you open the NodeODM interface you will be able to monitor and manage the task. This is also implemented in ClusterODM.
We have also replaced the .zip compression method in NodeODM to be faster, more memory efficient.
From the top of Whipps Ledges at Hinckley Reservation on November 16, 2016 (Kyle Lanzer/Cleveland Metroparks)
Reposted from smathermather.com
LiDAR and photogrammetric point clouds
If we want to understand terrain, we have a pricey solution and an inexpensive solution. For a pricey and well-loved solution, LiDAR is the tool of choice. It is synoptic, active (and therefore usable day or night), increasingly affordable (but still quite expensive), and works around even thick and tall evergreen vegetation (check out Oregon’s LiDAR specifications as compared with US federal ones, and you’ll understand that sometimes you have to turn the LiDAR all the way up to 11 to see through vegetation).
For a comparably affordable solution, photogrammetrically derived point clouds and the resultant elevation models like the ones we get from OpenDroneMap are sometimes an acceptable compromise. Yes, they don’t work well around vegetation in thickets and forests, and other continuous vegetation covers, but with a few hundred dollar drone, a decent camera, and a bit of field time, you can quickly collect some pretty cool datasets.
As it turns out, sometimes we can collect really great elevation datasets derived from photogrammetry under just the right conditions. More about that in a moment: first let’s talk a little about the locale:
Sharon Conglomerate and Whipps Ledges, Hinckley Reservation
One of my favorite rock formations in Northeast Ohio is Sharon Conglomerate. A mix of sandstone and proper conglomerate, Sharon is a stone in NEO that provides wonderful plant and animal habitats, and not coincidentally provides a source of coldwater springs, streams, and cool wetland habitats across the region. A quick but good overview of the geology of this formation can be found here:
One of the conglomerate outcrops in Cleveland Metroparks is Whipps Ledges in Hinckley Reservation. It’s a favorite NEO climbing location, great habitat, and a beautiful place to explore. We wanted to map it with a little more fidelity, so we did a flight in August hoping to see and map the rock formations in their glorious detail:
Unfortunately, as my geology friends and colleagues like to joke, to map out the conglomerate, we need to “scrape away the pesky green vegetation stuff first”. We don’t want to do this, of course — this is a cool ecological place because it’s a cool geological place! It just happens to be a very well vegetated rocky outcrop. The maple, beech, oak and other trees there take full advantage of the lovely water source the conglomerate provides, so we can’t even glean the benefits of mapping over sparse and lean xeric oak communities: this is a lush and verdant locale.
So yesterday, we flew Whipps Ledges again, but this time the leaves were off the trees. It can be a challenge still to get a good sense of the shape of the landform, even with leafless trees: forest floors do not provide good contrast with the trees above them, and it can be difficult to get good reconstructions of the terrain.
But yesterday, we were lucky: there was a thin layer of snow everywhere providing the needed contrast without being too thick to distort the height of the forest floor too much; shadows from the low sun created great textures on the featureless snow that could be used in matching.
The good, the bad, and the spectacular
So, how are the results? Let’s start with the bad. The orthophoto is a mess. There’s actually probably very little technically wrong with the orthophoto: the stitching is good, the continuity is excellent, the variation between scenes non-existent, the visual distortions minimal. But, it’s a bad orthophoto in that between the high contrast between the trees and the snow compounded with the shadows from the low, nearly cloudless sky result in a difficult to read and noisy orthophoto. Bad data for an orthophoto in; bad orthophoto out.
The orthophoto wasn’t our priority for these flights, however. We were aiming for good elevation models. How is our Digital Terrain Model (DTM)? It’s pretty good.
The DTM looks good on it’s own, and even compares quite favorably with a (admittedly dated, 2006) LiDAR dataset. It is crisp, shows the cliff features better than the LiDAR dataset, and represents the landform accurately:
So, if the ortho is bad and the DTM is good, what is great? The DSM is quite nice:
The DSM looks great. We get all the detail over the area of interest, each cliff face and boulder show up clearly in the escarpment.
Improvements in the next iteration
The digital surface model is really quite wonderful. In it we can see many of the major features of the formation, including named features like The Island, a clear delineation of the Main Wall and other features that don’t show in the existing terrain models.
Due to untuned filtering parameters, we filter out more of the features than we’d like in the terrain model itself. It would be nice to keep The Island and other smaller rocks that have separated from the primary escarpment. I expect that when we choose better parameters for deriving the terrain model from the surface model points, we can strike a good balance and get an even better terrain model.
Beating LiDAR at it’s own game
It is probably not fair to say we beat LiDAR at it’s own game. The LiDAR dataset we have to compare to is 13 years old, and a lot has improved in the intervening years. That said, with a $900 drone, free software, 35 minutes of flying, and two batteries, we reconstructed a better terrain model for this area than the professional version of 2006.
And we have control over all the final products. LiDAR filtering tends to remove features like this regardless of point density, because The Island and similar formations are difficult to distinguish in an automated fashion from buildings. Tune the model for one, and you remove the other.
For our use case, however, we can use the best parameters for this area, take a high touch approach, and create a really nice map of a special area in our parks for very low cost. High touch/low cost. I can’t think of a sweeter spot to reach.
I had an interesting question recently at a workshop: “What parameters do you use for OpenDroneMap?” Now, OpenDroneMap has a lot of configurability, lots of different parameters, and it can be difficult to sift through to find the right parameters for your dataset and use case. That said, the defaults tend to work pretty well for many projects, so I suspect (and hope) there are a lot of users who never have to worry much about these.
The easiest way to proceed, is to use some of the pre-built defaults in WebODM. These drop downs let you take advantage of the combination of a few different settings abstracted away for convenience, whether settings for processing Multispectral data, doing a Fast Orthophoto, flying over Buildings or Forest, etc.
You can also save your own custom settings. You will see at the bottom of this list “Steve’s Default”. This has a lot of the settings I commonly tweak from defaults.
Back to the question at hand: what parameters do I change and why? I’ll talk about 7 parameters that I regularly or occasionally change.
Occasionally we require a little more detail (sometimes we also want less!) in our 3D models from OpenDroneMap. Mesh octree depth is one of the parameters that helps control this. A higher number gives us higher detail. But, there are limits to what makes sense to set for this. I usually don’t go any higher than 11 or maybe 12.
Often with a dataset, I want to calculate a terrain model (DTM) or surface model (DSM) or both as part of the products. To ensure these calculate, we set the DTM and DSM flags. The larger category for DTM and DSM is Digital Elevation Model, or DEM. All flags that affect settings for both DTM and DSM are named accordingly.
OpenDroneMap often does a good job guessing what resolution our orthophoto and DEMs should be. But it can be useful to specify this and override the calculations if they aren’t correct. ignore-gsd is useful for this.
DEM Resolution applies to both DTMs and DSMs. A criterion that is useful to follow for this setting is 1/4th the orthophoto resolution. So, if you flew the orthophoto at a height that gives you 1cm resolution ortho imagery, your dem-resolution should probably be 4cm.
A related concept is depthmap resolution. Depthmaps can be thought of as little elevation models from the perspective of each of the image pairs. The resolution here is set in image space, not geographic coordinates. For Bayer style cameras (most cameras), aim for no more than 1/2 the linear resolution of the data. So if your data are 6000×4000 pixels, you don’t want a depthmap value greater than 3000.
That said, usually, 1/4 is a better, less noisy value, and depthmap calculations can be very computationally expensive. I rarely set this above 1024 pixels.
Camera Lens Type
I saved the best for last here. So, if you’ve made it this far in the blog post, this is the most important tip. In 2019, OpenSfM, our underlying Structure from Motion library, introduced the Brown-Conrady camera model as an option. The default for camera type is auto, which usually results in the use of a perspective camera, but Brown-Conrady is much better. Set your camera-lens to brown, and you will get much better results for most datasets. If it throws an error (which does happen with some images), just switch it back to auto and rerun. Brown will be a default in the near future.
(Reposted from https://smathermather.com/2019/12/02/self-calibration-of-cameras-from-drone-flights-part-3/)
I have been giving a lot of thought to sustainable ways to handle self calibration of cameras without undue additional time added to flights. For many projects, I have the luxury of spending a little more time to collect more data, but for larger projects, this isn’t a sustainable model. In a couple of previous posts (this one and this one), we started to address this question, pulling from the newly updated OpenDroneMap docs to highlight the recommendations there.
As I have been thinking about these recommendations, there are other more efficient ways to accomplish the same goal. Enter the calibration flight: the idea is that with some cadence, we have dedicated flights at the same height and flight speed as the larger flight in order to estimate lens distortion.
For this testing, I chose a relatively flat but slightly undulating area in Ohio in the USA: the Oak Openings region, which is a lake bottom clay lens overlayed with sand dunes from glacial lakes. It has enough topography to be interesting, but is flat enough to be sensitive to poor elevation models.
The test area flown is ~80 acres of residences, woodlots, and farmland.
Flown with a DJI Mavic Pro which has an uncalibrated lens with movable focus, the first question I wanted to address is how much distortion do we get in our resultant elevation models if we just allow for self calibration? It turns out, we get a lot:
We have seen this in other datasets, but this forms a good baseline for our subsequent work to remove this.
Next step, we fly a calibration pattern. In this case, I plotted an area large enough to capture two passes of data, plus an orbit around the exterior of the area with the camera angled at 45° for a total of 3 minutes and 20 seconds.
When we process this data in OpenDroneMap, we can extract the cameras.json file (either in the processing directory or we can download from WebODM) and use that in another model. We can do this using the cameras parameter on the command line or in WebODM through uploading the json file from our calibration dataset.
But, before we do that, let’s do a review of our calibration data — process it and take a look at what kind of output we get. First, we process it using defaults and evaluate the elevation model to look for artifacts that might indicated whether the calibration pattern wasn’t successful.
Our terrain model from the Ohio Statewide Imagery Program elevation model looks like this for our calibration area:
Note that this is mostly a moderately flat farm field with a road and small ditches running North/South in the west of the image and a deep Northeast Ohio Classic ditch in the east.
How does our data from our calibration flight look?
It’s not bad. We can see the basic structure of the landscape — from the road in the west to the gentle drop in elevation in the east.
Our default camera model is a perspective camera. How does this look with the Brown–Conrady camera model that Mapillary recently introduced into OpenSfM?
With the Brown–Conrady camera model, we see additional definition of the road bed, ditches alongside the road, and even furroughs that have been ploughed into the field. For this small area, it appears the Brown–Conrady camera model is really improving our overall rendering of the digital terrain model, likely as a result of an improved structure from motion product. We even see the small rise in the field at the southern central part of the study area, and as with the default (perspective) model, the slope down toward the ditch on the east of the study area.
With running these both with perspective and Brown–Conrady cameras, we can apply those camera models as fixed parameters for our larger area and see what kind of results we get.
Our absolute values aren’t correct (which we expect), but the relative shape is — the dataset is now appropriately relatively flat with clear delineation of some of the sand features. This is the goal, and we have achieved it with some of the most challenging data.
How does our Brown–Conrady calibration model turn out? It did so well on the small scale, will we see similar results over the larger area?
In this case, no: the Brown–Conrady model over compensates for our distortion parameters. More tests need to be done in order to understand why. For now, I recommend using the perspective model for corrections on large datasets, and Brown–Conrady camera model on smaller datasets where the details matter, but the distortion isn’t discernible.