I’m new to this. I recently completed a job at a refinery and when I went to upload the images, my laptop died before it completed. I took a total of 500 images, but the computer died after only uploading 394. I was unable to charge it and resume, so DD I guess thought I was done and processed what I had uploaded. When looking at the map showing the location of the photos, the entire middle section was missing, so I only had a left and a right section basically.
When the map was finished, I took a look and on the 2D image, there are some very rough spots up the middle section, which I wasn’t suprised by since it missed those photos on the upload, but the 3D map looked pretty good and was basically acceptable. I figured both would improve when I uploaded all 500 photos and I thought the 3D map would be damn near perfect.
The weird thing is the 3D map with all 500 photos looks way worse. There is a large square structure in the middle and it looks like a comet hit it because a huge section is missing from the side/middle, but also overall the 3D map just looks inferior than the prior version done with fewer photos.
The photos I uploaded were about 475 top down, and about 25 ortho, mostly of the large structure that looks like crap. Any idea as to why this happened? Could it be a processing issue? Should I reupload the images and try again?
It’s interesting you point this out as I regularly have the same problem and generally have to upload the images using a variety of different permitations before I get it right. Sometimes I give up on the 3D and just settle with a good 2d map. Sometimes I get better results with just nadir and no oblique, sometimes it’s better with one crosshatch mission added, other times with 2 crosshatch. On one occasion I uploaded the nadir and added half a dozen oblique. That was probably my second best 3D result and have no explanation for it.
The best model I produced was at 60m high, sunny day, one nadir mission using manual exposure and one crosshatch mission using automatic exposure. I thought with a variety of under and over exposed images I was going to be presented with an awful process but to the contrary it was the best.
Other than raise the altitude by 10m from my usual 50m I hadn’t done anything differently so wondered if raising it and giving the camera a wider field of view was the answer. With that I raised it again to 70m on my last and results were disappointing.
What else I have learnt is that I have also received very different results from uploading the same images and is possibly what frustrates me the most. If I knew it was me that was the problem I could fix it, but when I’m receiving varied results using the exact same images, it’s out of my hands. My unqualified theory is that it depends what day of the week it is and what level of traffic is passing through the servers at the time. Upload at a quiet time and it might take a little longer to think about it but being an international resource, I haven’t yet determined when is best to begin the process.
Thanks for the response. I read elsewhere that including obliques when you only want a 2D map can result in worse results. There is a thread where the user compares the 2D output with obliques and without and the map without is way better than the map with. I’m going to try that. I’ve reuploaded using only nadir images and I’m going to see if that improves the 2D map. I’m also interested to see how that affects the 3D map.
I guess I’ll give a shot at reuploading all of the images too and see if I can get a better result. I was really optimistic when I saw the 3D map based on only 394 images… I just assumed the finished result would be even better, so I was suprised when it actually looked worse.
Sculpting good 3D models with a drone is part art and part science at this time. Gaining experience in how best to apply the techniques to optimize the particular model is the best approach.
Along with the Nadir grid photos, getting good sequential overlapping oblique images without the sky in them is key. Many people use the POI feature in the DJI Go app to get good overlapping oblique images or the Structure Mode during the flight. Sometimes nadir images alone work better, sometimes both nadir and oblique, and sometimes obliques alone (when done properly) work better.
These documents give good insight on how to go about it.
Unfortunately Gary I have read your links over and over each time you post them when another member says they too are having trouble. As per my previous email, adding two 70% overlap crosshatch missions including the best of 6-800 oblique images to 900 x 85% nadir images (bearing in mind my focus is on built up landscapes rather than single buildings) does not guarantee me good results first time around (or even close) and I don’t understand why. Quite often I can receive better results with no oblique images or just half a dozen. More often than not if I run the process 3 or 4 times using the same images uploaded, one of those four attempts will give me a good result. Again, I don’t know why as all I’ve done is run the same process over and over using the same images. No fewer, no more.
If I run a mission or take photos using varying heights without the sky, the elevational results of the process are disastrous. What I thought would be clever would be to run the nadir at x altitude and a crosshatch at half of that altitude with a raised gimbal, again without sky. The worst result I think I ever received.
It’s annoying because there’s only so many ways and times I can run it and when even the same uploads don’t give the same results it makes me wonder whether I have much influence at all. Rather than science or art I’m beginning to think it’s fluke if I get a good one. The only thing I can usually rely on is a good orthometric result using nadir and crosshatch. 3D is a bonus but I really do wish I could work out the secret formula as I’m sure it’s something at your end that determines what kind of result I’ll get back next time. It has to be if I’m running identical processes and receiving different results each time.
I have never seen quite a variation from performing different types of flights. Yes ground or 3D is better with one or the other, but not as much as is being described here. As a default for a structure model I run the traditional Nadir and then try to get 3-5 varied angles of each structure face at the same elevation. I.E. Looking 30d at the center of the face, 45, 90, -45 and -30. You have to have enough images per face to stitch nicely, but I have also experienced that too many images can distort or cause a ribboning effect on the render of the face. A dirty illustration, but the point is that I am catching 3-5 shots/face while maintaining the same distance/height from the building. Structure mode does a pretty good job of this if you offset the perimeter of the building equally, but in my opinion it takes too many shots and sometimes doesn’t capture exactly the right ones. I would prefer a method that flies the pattern, but let’s me take the pictures. A plus to this is that not only am I capturing the obliques, but nice progress photos at the same time.
It’s a shame, because I only had a tidy up of my desktop last week and deleted all of the failed attempts. Daft thing is I thought at the time whilst I was deleting them that I’m going to need these to prove my claim. Never mind, I’m sure there’s more to come.
Structure models only allow me 500 images so I’m left with having to opt for Terrain when processing, even though my subject includes as much structure as it does terrain. This is something I could experiment with on a one-off small area though, just to see how much a difference Structure makes over Terrain. On the model we discussed several months ago @MichaelL, I think from memory that was before Crosshatch was released so I did automate around a POI at various heights after the nadir mission to no avail. In fact, that’s the one provided me with a 30m high mountainous range across the middle of a flat field.
Sorry to hear this James.
"Processing maps relies on stochastic (randomised) methods of solving complex optimization problems. As a result, running the same images through the system twice will always give you slightly different results. Usually these differences end up being very minor, but in areas with low image overlap, the algorithms are stretched to their limit and more differences may appear.
You can usually reduce these effects by flying higher above the target, or by increasing overlap when flying over hills, tall trees and buildings."
Each subject matter is unique and may require a unique application of the technique to optimize it. That is where experience is key. You are gaining that, but that takes time unfortunately. I rarely rely upon autonomous flights to optimize the particular features of the subject matter. They may be part of it, but it usually takes custom, often manual flights to optimize the particular details.
I did this one by flying manual orbits with the DJI GO4 app at two altitudes shooting 85 overlapping oblique images looking down 45 – 60 degrees. I did not shoot any vertical grid images.
Gary, do you have any other building models flown in the same manner? I know you are proud of this one, but I flew 4 different structures just last week that all had their own characteristics. This method may have worked for a simple rectangular building, but try something a little more complicated. I like your “part art” tag because accounting for all these variables is just that. I have even had to adjust my flight for proper angles because the sun wasn’t in the right spot to capture the part of the structure we wanted.
I agree with you Michael. Sculpting 3D models with a drone is part art and part science. Getting the science optimized to work for you is the art. Each subject matter is unique and requires adapting the techniques to optimize the features and results.
Here is a bigger building on our blog.
