Image Intensive: Suggestions for improving model accuracy

I wanted to start a general thread to get user input and suggestions on improving (vertical) quality of surveys. By way of example this is a dam survey with a combination of nadir and oblique imagery…

Note loss of data on top of dam…

Vertical data lost on treatment tanks…

Loss of information on chemical towers (a separate oblique orbit was run around these).

The weeping willow effect…

There are also a number of other distortions that can be seen in these shots, which were a combination of 223 nadir and 221 oblique images covering just under 50 acres. Oblique images included orbits of dam, parallel close-ups of dam face, and separate orbits of chemical towers, main buildings and treatment tanks. Each was a single pass at approx 45 degree camera angle.

I am interested to hear what methods and suggestions people might have to improve the quality and vertical accuracy of these kinds of surveys, with particular reference to the DJI / DroneDeploy platform? Whilst there are many papers on this kind of thing, different platforms behave a little differently and so I wanted to find out what people had found worked and what hadn’t.

I would imagine that there is potential scope for at least some of these methods to be incorporated into flight planning in the future.

[Whilst the actual deliverables in this project did not require an accurate 3D model it is a good case in point for the above]

Hi Roger,

This is an important topic to be discussed.
It would be very interesting if the DD could improve the quality of 3D.
Your oblique images were all taken at 45 degrees?

Did you run these using the 3d Model setting when uploading? That will reduce or remove the “draping” effect.

@chasemgray Yep. I ran this with Map Engine / Structures. It was less a question of the DD engine and more a question of improvements to technique, really, as I get similar results on other engines. I notice other people’s models show similar issues so I thought it worth opening it up to the wider audience on the forum for comment.

@Roger_Olmos Yes, the camera was at 45° although due to flight and landscape characteristics a number of those frames are not ideal - contain a fair amount of sky / water etc, hence my comment about better pre-selection. The chemical towers were a bit of a surprise as these comprise of a close-up orbit, a distance-orbit at different height and nadir fly-over. I’m going to model those frames separately on another engine to figure out what is wrong with them.

Following on from a similar discussion here I removed any images where there was horizon in the shot and reprocessed.

The chemical towers are much crisper, but still lacking a little higher vertical detail…

The flight path for these looks like this…

And the individual renderings in Pix4D look like…




The top of the dam is much improved…

The treatment tanks have not benefited much. I suspect that these need more oblique to get the detail. That gantry looks pretty wavy for a straight lump of metal.

Again, the flight path, which you will see has an incomplete orbit. That is because some of the images were pulled from the set.

and the Pix4D…

Generally, though, pre-sorting the imagery to pull out anything <45° or containing (significant) horizon has improved the output quality of the model a fair amount.


Thanks for the detailed notes and screenshots - all really helpful to us as we’re constantly working on improving the Map Engine. Great to hear you’re having more success with the 45deg obliques, and the dam model you’ve linked looks fantastic!

For this project I noticed you selected the terrain option for the model when you uploaded the images. This assumes that the scene is largely flat (e.g. a field); if you have scenery with a lot of vertical detail such as the treatment tanks and chemical towers in your example then I’d recommend trying out the “structures” option when uploading as this is optimised for 3D scenery / vertical structures. N.b. this option will take a bit longer to process, especially for large image sets. Maybe you could try re-uploading the same data set with the structure option for a comparison?