Best photography plan for a building site with trees?

[ThursdayJuly23rd]

Thanks for any suggestions on how to best go about photographing this building area. See screenshots below for more details and proposed photo paths.

[cam3d]

Hi ThursdayJuly23rd -

From the provided information I'd recommend looping around each tree and cabin etc as this will give you the greatest coverage of the scene and you're unlikely to miss any parts of it - Don't forget to shoot the ground between subject loops otherwise you might end up with a bunch of disconnected components -

As you mentioned the trees are pretty dense it sounds like you will be taking photographs in a low light environment - I highly recommend using a tripod to allow slower shutter speeds and low ISO for the best image quality your camera can provide.

[ThursdayJuly23rd]

Thank you. For a tree/dirt/grass subject like this do you still recommend 70 to 80% overlap or would you increase to 80-90% overlap? With regards to "shooting the ground" ... should I also be doing loops around these open ground areas? Something like this:



Here is a blank version in case you want to illustrate your own Thanks very much!

[ThursdayJuly23rd]

My 2nd visit to this site produced bad results again even though I thought the lighting conditions, camera settings, and photo path were improved.

[DanielMuirhead]

Hi, I am late to the party here, but in case it is relevant I just finished a 180k photo capture (and 35k photogrammetry model, so far) of a graveyard: https://skfb.ly/6UYNO

[ThursdayJuly23rd]

Daniel; thanks for your helpful recommendations. I enjoyed reading your post about the Howff project. Seems like you have created a mini version of the google car . It looks like you just walked a "grid" pattern with your custom 3D camera and alternated N S E W and NE SW NW SE. This looks like it did a perfect job of covering the grounds. Though is seems like you wouldn't have 60% overlap between each photo. Maybe the volume of photos you took makes up for that?

On the two previous visits to the site I'm trying to model I did not use a grid pattern. It was recommended to me to orbit around each tree and then orbit around the open/clear areas. This has led to mixed results; though possibly more the fault of the photography settings and number of photos I captured.

Here you can see the result of my 2nd attempt at photographing the scene.... http://somup.com/cYQFVuXmrW . Big issues with the open ground area. This open ground area has several survey stakes protruding from the ground and I'd like those to be captured accurately as the home builders will use those points when proposing home designs.

With regards to your advice on how to best capture the ground.... please see screenshot:

[DanielMuirhead]

Thank you for sharing the results of the 704 photo capture, that was informative to see the results.

First off, cam3d knows what they are talking about so I am not going to contradict any of their recommendations.

As I understand it, if a 360d camera produces photos with a resolution of 5376 (horizontal) x 2688 (vertical), then that is 5376px per 360d field of view (FoV). So for a FoV similar to your DSLR (assuming 18mm lens = 90d horizontal FoV, going by this random table I just internet-searched, https://www.nikonians.org/reviews/fov-tables), the equivalent photo 'horizontal' resolution (post-decomposition) would be 5376/4=1344px, compared against your DSLR which (I think you said) is generating photos with at least 3k resolution. So that low resolution issue is how I would be quantifying the quality of 360d versus the 18mm/DSLR. And if I was to engage in pure speculation (as I know nothing much about 360d cameras, and I am happy to be corrected on this) I would imagine that the sensor used in the 360d camera will be inferior to that used in the DSLR. By analogy, if I capture a 2,000x1,000px image on a DSLR, then capture the same resolution image on a cheap smartphone or cheap compact camera, the DSLR image will contain more useful data due to the high quality sensor (again, my assumption, am happy to be corrected on that by someone with superior knowledge).

In regard to percentage overlap between photos, e.g. 60% or 80% or such, I do not really calculate that in advance when performing photo capture. I am probably performing some instinctive calculation at some level (I have captured around 300k photos with a regular camera, of photos for use in photogrammetry files, so have gained some experience), but am not consciously reflecting on that when in the field. In regard to my graveyard capture, for example, there would be some way of formally quantifying the extent of overlap - e.g. calculations involving distance of camera from subject, distance between respective capture locations - but my first field-test of the rig, 20 minutes in duration, was enough to show that the capture method (and photogrammetry reconstruction) was successful, so after establishing that I just bashed on and performed 12 hours of subsequent capture, in the knowledge that there should be some useable data gained. Plus my habit was to err on the side of 'gather too much data', by only taking small steps between capture positions, instead of taking large strides between positions - although on some occasions I did intentionally take large strides/infrequent polling to gauge the effect of in/frequent polling on quality of reconstruction.

Part of what informed my rig design was my (heretofore) capture habits for street length scenes. My habit is to capture the same part of the building facade from three angles, being 1) perpendicular, 2) at 45d from one side, 3) at 45d from the other side. By example, this photo set of a wall: https://www.flickr.com/photos/daniel...57715977818911. Excepting the first photos (the wall end) where a few one-off photos are taken to ease into the subject, you will notice that the photos are captured in groups of 3, 1) facing left, 2) straight on, 3) facing right. By using this capture pattern for the whole length, I will eventually capture the one part of facade from three angles (front, left, right), but I do this via capturing all 3 photos at once.

Similarly, working across the entire graveyard area with my camera rig, I am going to end up capturing the one piece of ground from 8 different angles (reg/60d cameras at N/E/etc. then at NE/SE/etc.), and arguably from 16 different angles if we include the wide/120d cameras.

Bearing all this in mind - and good work on reading this far! - my recommendation in regard to how to proceed with the next photo capture of your site, (if you have time for at least 1 session of tests and 1 of final capture) would be to capture a few test sets: 1) as close to perpendicular to the ground as you can manage, achieved by holding the DSLR above your head and oriented down to the ground, probably angled in such a way that you are not in frame, performing blind capture (capture a few photos, see how they turn out, then proceed with capture); 2) a similar set but at 45d to ground. The flythrough you uploaded showed that the model had some errors/anomalies in regard to the ground, namely the ground dropped away while the trees were mostly reproduced okay. I would hypothesise that is because the cameras/photos were all captured perpendicular to the trees and other upright items, whereas those same photos were capturing the ground only at a glancing angle. I get the exact same errors/anomalies when performing street length capture, i.e. the ground dropping away, because my photos are angled perpendicular to the building facades but at only a glancing angle to the ground, so the road/pavement gets erroneously reconstructed. It is mainly why, for my graveyard capture, I was explicitly aiming my cameras down towards the ground, to get a good ground model.

Depending on your DSLR, it should be possible to engage in fairly rapid capture. When I am capturing, I am taking as many photos as I can in as short a time as possible (while trying to keep the camera still to reduce blurriness/motion blur etc.). So e.g. with my DSLR I can either capture via looking at the preview through the integrated mini screen and observe the result after every photo (which takes at least 20s per photo) or I can turn off the screen and instead use the viewfinder or just perform blind capture and not bother checking the results before taking the next shot (which brings the capture time down to 2s per photo). When performing urban capture, street lengths and such, I sustain a rate of around 2k DSLR photos per hour (single handheld camera), which often comes down to 1k/hour due to temporary obstructions (people, cyclists, cars, buses).

That would be super if you could operate the additional cameras in such a way that increases your overall capture rate. Another two things I would mention is: if there was some way to set up your 360d camera to passively capture while you were traversing around the area, e.g. if you could have it continuously videoing or timelapse-photographing, sticking out a rucksack on your back, then that would be one way to capture secondary images that you could experiment with. Secondly, video would be another way to perform rapid capture of low-quality/preview data. Video data is inferior to photo data (although a high quality camcorder will give superior results versus a low quality still camera etc.) but if you could capture a walkthrough of the whole area on e.g. your smartphone, with the camera pointing down at ground, slowly pacing across the area in a grid trajectory - that could be an additional data set to experiment with.

Anyhoo, all the best with your subsequent photo capture, and let us know how you get on!

*Edit: it is good that you are persisting, I feel your pain, but we will get there eventually! These things can be frustrating. I sometimes spend a load of hours fruitlessly trying to get something to work, only to slightly adjust my approach and solve the whole thing. So keep up the good work, the desired results will come!

2nd Edit: in regard to your photo capture of the site, if you proceed with photo capturing the ground in the way I mention above (camera raised above head and angled down towards ground), then I forgot to recommend, you could follow an approximate grid trajectory that would be something like: take photo1 while facing N, sidestep then orient NE to take photo2, sidestep then back to N for photo3, etc. Then, once you reach the other end of that trajectory, start working back while doing the same thing, face S and capture photo101, sidestep and face SE then capture photo 102, etc. Do not worry about retracing your steps re exact capture positions, having minor variations in camera position and orientation is actually advantageous. All the best!

[ThursdayJuly23rd]

Incredibly helpful. Thank you so much for the information rich and detailed response. I have a black magic cinema camera 6K that might come in handy for the video coverage. Maybe I will create a separate post to discuss best video capture techniques. Thank you again. I will report back with my progress :-)

[ThursdayJuly23rd]

Daniel; your technique worked great on my last test visit ( though it looks like I'll need to capture more shots facing level and above to better capture the full tree trunk and upper branches ) ... please see this video for details on how it turned out: http://somup.com/cYQqFdXdTw

I captured the images using my bmpcc6k camera on video mode shooting one frame every three seconds. I created a post about the video to image process here. I'm gonna try to set my gimbal up on an automated repeatable pattern to capture the frames. Hopefully that will be successfull; I'll report back with my findings

Here are some screenshots from the model showing good detail in the ground.



Here is the capture pattern I used... as you can see the ground came out good but the trees did not get captured above the lower stump areas. Think this is because I didn't get enough shots of the trees at upward angles?

[DanielMuirhead]

Hi ThursdayJuly23rd, apologies for my late response, I am busy with some stuff right now. It is good that you are documenting your process through, including through the video walkthroughs, they are informative to watch.

That is fantastic to hear about your progress, great work! As a quick answer to the tree trunk issue, yes I would imagine that the limited reconstruction of the trunks upwards could follow from a limited capture of that part of the subject. Now that you have a capable method for doing the whole ground area, it would now be possible to focus on parts of that ground, e.g. individual trees. Edit*: this in the sense that it would now be possible to focus on capturing individual trees with a view to integrating those new photos into that ground-set you have already established. Either through running the whole SfM/Sparse again, or through using 3DF's 'Merge Workspace' to e.g. process each tree individually (SfM/Sparse stage) then merge it into the ground model (for Dense stage onwards).

One option could be a capture of each tree, the camera located at 1m or 2m from the trunk centre, orbiting around that centre, capturing a photo/frame at every 5d or 10d. I think maybe you already performed an orbit capture for each tree during a previous capture session? So it might even be possible to integrate those photos with that video/frame set shown above. And/or starting afresh, begin the orbit capture at the tree trunk base, with the camera angled down at the transition area between ground-roots, then gradually orbiting upwards, altering the pitch of the camera accordingly, while maintaining high overlap between both successive photos and successive orbits, and working in a 'corkscrew' trajectory up the height of the tree, capturing as high as you intend to reconstruct (although bearing in mind the reconstruction might get a bit patchy when the tree transitions from trunk-cylindrical to branches-complex).

I am going to check back in soon, but it looks like you are heading in the right direction, so great work in persisting!

*Edited 2nd time to improve text clarity.

[ThursdayJuly23rd]

My latest scan attempt went well ( although I'm now having issues with my pc crashing when building the Mesh on 1449 images ) but I'm hoping to increase my capture routine efficiency further by capturing less images when not close to trees or other objects other than the ground. I captured about 70 images of each position ( I haven't tested my automated gimbal plan yet ) . It was pretty fast to capture the images by hand at each position using my video camera to capture a frame every two seconds. So each position took about 2.5 minutes.

Here is a video showing my latest results... 19 scan points 1449 images.... thanks for any input on it.

My question is:
In areas where I'm not close to a tree or house and I just need to capture the ground; do I need to capture images facing outwards and upwards or just the ground? Below are two figures to better illustrate my question.


It seems like in Figure B the full dome images not close to the trees would just have redundant information.

[DanielMuirhead]

Hi ThursdayJuly23rd, it is fantastic to read about your progress. Just a heads-up that I am stepping back from the forums so I am no longer going to be active here, but feel free to contact me through the email listed on my personal website if I can provide any more advice. All the best!

[Andrea Alessi]

It seems you are using a nodal point method of scanning, which won't guarantee enough parallax for quality results. Rather than staying still and rotating, please rotate around your subject.

You can find a (crudely) drawn example in your other thread. Do you want me to merge the threads? I missed this one

https://www.3dflow.net/forums/forum/3df-zephyr-forum-english/6652-why-images-are-not-aligning-when-grouped-with-other-photos-from-same-photoshoot