Totally new territory to me … Perhaps I’m asking for too much, but I’m somehow disappointed by the quality of my first 2D map ever, created with default settings and DJI P4P drone. When magnified, many details appear badly deformed, smudged like. I did my homework with DroneDeploy tutorials, I’m a seasoned aerial photographer, familiar with image stitching technology and requirements for successful mapping imagery. I don’t believe that I can improve the quality of images delivered for processing, although I may be wrong. I tend to suspect that the 3D option is somehow messing with the quality of 2D stitching. I’m not much into 3D, all I need is a detailed 2D map utilizing the resolution of provided images to a possible level of the truth. Is there any way to improve just that at the expense of 3D greatly overrated significance?
Thanks for swinging by the DroneDeploy Forum. Can you share some sample images of your map so we have a better idea of what the issue is?
Also, can you verify that you have “Set Exposure Manually in DJI Go” and “Set Focus Manually in DJI Go” toggled off in your Advanced settings of your flight plan?
Keep me posted,
Christina, thank you for response. You should be able to see the artifacts on first image, next is the original image.
However, I’m a bit confused by these advanced camera settings requirements … If it’s toggled off, than what (or who?) will take care of these critical set ups?
You might try the crosshatch flight pattern in this app to get perspective images from two angles to help alleviate issues like this. The altitude you fly based on the subject matter may also affect this. Some of it comes down to getting experience on how to optimize the results.
Oblique images also can help but adapting them to optimize the subject matter is sometimes necessary.
The perimeter Structure Mode may help with this by automatically getting some obliques.
Gary, thank you for advice. Although I’m a seasoned aerial photographer, mapping is a totally new territory to me. The goal is to create 2D map of a small town with as much detail as possible, to replace government’s satellite image of poor resolution. Therefore I may ignore 3D issues completely, focusing on stitching/warping artifacts instead. Perhaps higher attitude will help to reduce these deformations, naturally with affordable loss of resolution?
It’s interesting you raise this @Matthew. I’ve been flying the same way for just under a year now but recently have noticed a lot of this warping. I’m not sure what’s causing it.
In addition to my standard Structures inclusion, I flew for the first time with the crosshatch this week @Gary mentioned to try to irradiate it and incorporated the images into the vertical/structures upload. The overlap of the crosshatch was something like 75% - the results were horrendous. I removed the crosshatch and results were better. I added a few crosshatch images back and the map was one of the best I’ve produced and the process was several hours long. Tick box.
Having considered the crosshatch overlap may have been too high previously, I used the default 60% today, incorporated them into the verticals and the results were still not great. Actually, they’re worse than not great and it’s frustrating because I’m not flying any different to how I’ve been flying for a year.
If I run the process again I will almost guarantee I’ll receive different results again, but I don’t know why.
Something I have been noticing over the past few weeks is that process times are wildly different between different attempts. This might be a mad thing to consider but fewer images seem to be taking a longer time to process but returning better results. Something’s changed in the process and it’s not for the better. It’s almost as if the system has been tweaked to provide quicker results and take the strain off the heavy workload as DD takes on more users, but the results are suffering as a result. I don’t know. I can’t put my finger on it. It seems to be pot luck now whether I receive a good process or not.
Probably not making a lot of sense but I know what I mean. Something’s changed.
Ran the process again with no changes to the images covering this area of the map…
Interestingly the process has opted to use different images to create the finished map as can be seen by way of the parked car changing in the street. But as can also be seen from the outline of the grey roofed building, no blurred edges and no screwed up blue van and black car. The exact same images were uploaded for this area of the map. I haven’t corrected this. The third process did.
There is definitely an art to using this software. I thought I’d mastered it but not quite.
If I understand correctly you didn’t noticed such “smudges” before. That I can’t confirm since I just arrived on this territory. All I’m saying is that I’m not impressed much by the quality of my first attempts. The map looks OK at first glance until you try to zoom in to take the advantage of high resolution and see all these ugly distortions, warps, exploded walls, dented vehicles etc. I have no idea what to think about …
So I processed another map and the results again were initially disastrous. The first mission was run at 50m high and had 85% sidelap / 75% frontlap with ‘Structures’ selected. It took a total of 280 images, 168 of which were Nadir, 112 Structures oblique.
A second 50m mission was flown using ‘Crosshatch’ at 60 side/frontlap and the gimbal set at 65 deg. pitch to avoid the horizon. This produced a further 88 oblique images at approximately 45 deg. to the previous Nadir flight.
All images were in focus and look good having manually set through Go4.
For the first process I quite reasonably uploaded all 368 images and this was the result. [All maps below have been calibrated to the same point]
I’ve found it easier to detect the clarity and sharpness of the map by using the Elevation view and you can easily see that there is barely anything square about the square buildings I have surveyed and the maps looks like something that wouldn’t be out of place in Gummy Land.
Because of the luck I’ve had recently I tried the process again, using different variations of the images. The second attempt was using just the 280 images taken in the first flight ie. the original Nadir/Structures. Results again were just as bad (blobby and jelly like) although it’s colour default was massively different from the first - recording the map’s height extents dramatically differently. See below:-
The third attempt below processed concurrently with the second, uploading only the Nadir images from the first flight and crosshatch images from the second ie. 256 images.
Now the first map took just 2.5 hours to process whilst the other two took about 7 hours. Of course that could be down to an increase of traffic but does appear to me that my more successful maps have taken longer to process and I wonder if the system is merely taking longer to think about it.
I may run a few other maps with different variations of image uploads to see if I can find a consistenct way to produce the optimum result.
James, it seems we’re on the same crusade, although I’m with much lower rank. I shoot my hood yesterday with roughly identical approach of your first flight in Structural mode, and get back equally smeared, rather useless render as on your 2nd sample. Shooting two sequences (nadirs + crosshatch) seems a bit troublesome, but perhaps this is the way to go for projects of critical importance. Would you mind to guide me how to execute second flight with crosshatched pattern for Phantom 4 Pro)?
BTW, it’ll be also interesting to find out if repeated downloads of identical material (re-named JPG files?) will render different results. Let’s stay in touch …
Honestly, I just want it to work consistently. Using a primitive Mavic I’ve come to master the focusing issue it’s plagued with under DD but that’s fine. It’s part and parcel of the course and when I upgrade to a better drone I’ll have something extra to look forward to.
It’s because I think I’m doing everything right and making the finer adjustments necessary to produce a good map, that it becomes frustrating knowing you did everything as you should but yet without warning three consecutive maps are rubbish on the first process and you have to experiment further to work out what random combination is required next.
That said, if I have to carry out a Nadir mission followed by a Crosshatch immediately after, so be it - as long as the goal posts don’t move again.
Assuming your P4 @Matthew operates as my Mavic, set up a standard flight in the normal way and when ready to go, copy the map (duplicating the area) and swap the flight type to Crosshatch at the bottom of the menu. If the Structures important to your final map are close to the edge of the original boundary line, extend them out by a further 20m or so (depending on height) to ensure those same structures are captured in the oblique 65deg.(?) images.
Once you’ve run the two missions and have two sets of images, upload them all into one single mission and see what comes out the other end.
James, pardon my persistence … I don’t see Crosshatch option anywhere … Do you simply set the flight direction at 45 deg and gimbal angle at approx. -65 deg manually for second flight? Both flights are in plain mode, no Structural?
Not at all.
The Crosshatch actually forms part of an app found in the App Market
Once installed it gives you additional flight options at the bottom of the mission setup.
The first time I used it I increased the overlap as normal and ran the paths parallel and perpendicular to the original Nadir path. I don’t know if it was the sheer number of oblique images or the fact they were taken at the 0/90 degrees that made it give rubbish results but the second time I tried it (set of three images above) I ran the Crosshatch at 45 deg. to the original Nadir and abandoned the Structures photos. The gimbal angle I left at 65 deg. and the overlap at 60% in both directions.
Thank you very much, James.
Just a tip to get your structures the best you can try to run parallel and perpendicular to the primary faces of the buildings regardless of which direction you’re nadir flight is running.
You mean parallel/perpendicular to the “primary faces of the buildings”, right? That sounds like “mission impossible” for larger area to be mapped, IMHO. 3D structural mapping isn’t my goal at all, but it seems to be somehow related to the quality of 2D map. Thanks for the tip anyway.
Yes faces, I have corrected it. The crosshatch is only needed at the buildings. Close to 75% of our work is civil so the majority of my missions are nadir only. Oblique imagery does not help the accuracy of the model. Especially considering that we are creating a DTM from that. It will be nice when this fact is well-known because this discussion happens far too often. I’ve done extensive testing by running both types of missions on civil sites. If you have access to survey gear you might want to do this yourself. Make a mesh from the Point Cloud that can be loaded into that survey gear and steak the surface. You can switch between surfaces at various points to verify it for yourself.
Will try that again also in that case. Oblique images taken perpendicular to the faces for me failed miserably last time. Running them at 45 deg. worked better.
Maybe I’ll run at 45 and 90 deg Crosshatch and experiment a bit.
Like @Matthew, primarily all I’m after is a sharp 2d map with relatively okayish height information. Detailed 3D facades would be a bonus but not essential.
Gentleman, at this moment of discussion I tend to believe that my expectations are simply too high. The images taken with usually wide angle lenses contains too much of oblique’d information to be truthfully warped and stitched into planar 2D projection. While this technology may produce stunning 2D/3D results when mapping (scanning?) smaller objects like a single worksite or real estate, it fails to deliver un-warped details of 50-100 acre developed area. High attitude and longer lens should provide much less distorted nadirs, but the amount of images to be taken for large map will challenge drone’s capability. Please correct me if I’m wrong …
I totally get what you’re saying about the warping from the camera, but our construction projects are on average right in that range of 75 acres and I haven’t experienced any issues with accuracy yet. I wonder what it would be like if the warping didn’t occur?!
Michael, map accuracy is not the issue here nor is questioned. Artifacts visible on 2D projection, however, are the reason for this thread. Seemingly some data responsible for 3D projection is leaking into 2D map, badly deforming some details in result. We are trying to determine the best possible approach to eliminate such deformation to acceptable level and without getting crazy. So far no luck on my side …