“I thought you were gone for good… :-)”
Nevah! Yes, been ‘working’.
Went to Bentley free photogrammetry seminar London yesterday - second annual one, top industry guys there, very insightful. Learned a great-sounding formula/system for interior photo-taking - will give it a try and feed it in. All advice adds more photos to the mix - I still hope to find the reliable precision/simplified-set.
Can you please share this great sounding formula ?? !! =) plz
Hi Tom, good to know!
BTW, the RC people are a bit funny about competitor names, so it would be better to avoid naming them directly.
Looking at the photos II’m not sure if it can be achieved withouth manual intervention at all. The repeating pattern on the wll is basically the only suitable surface, the rest varies between bad and impossible. So most if not all of the features will be detected on those walls which will inevitably lead to a false alignment.
Maybe with a ton of CPs you might be able to coax RC into reconsidering…
And you definitely need to mask the mirror - that’s the final straw to entirely confuse RC.
Also, you could use the tile pattern to establish a grid of GCPs - you don’t even need the real measurements, just the fact that they are square should suffice to do it.
About the mirrors… maybe not totally applicable here (I didn’t look at the pictures in details), but I had very interesting results with mirrors : actually when the space that is “inside the mirror” does not collide with the real space it can help aligment : let’s pretend this is just a window to another room with - what a coïncidence - an identical content. This will be reconstructed by RC, just as if this is a window. I even had interesting results inverting the pictures left-right and aligning them together : RC is tricked, space is reversed and this can help reconstructing parts of objects that are close to the mirror…
Of course, if the mirror is inside the room (and not on a side == wall), this is another story and RC cuts holes in the space trying to resolve 2 incompatible constraints : the real space and the space inside the mirror. And when you have more than one mirror this is even something else.
So I would not mask the mirror - but now I want to try this dataset too - looks like a small contest : who will get the best out of this difficult dataset ? 
Here is my 5 cents :
instead of trying to align everything together (that gives false alignments) I would do small alignments by groups of pictures where you carefully try to avoid false positives (group by proximity which should be easy because you shot in sequence moving a little bit) - then you export the components and align together. I am going to give it a try.
Jonathan,
I was thinking the same thing
That’s a great idea… I’ll give that a shot today as well.
For what it’s worth, I also wanted to suggest doing it for example in individual walls.
Jonathan, sounds interesting, the reverse engineering - did you then just crop the mirror part? How would that influence image geometry?
I thought of using a mirror to reach difficult places…
What I also wanted to say: I think in this case masking the mirror is worth a try since it will cut down on the confusable surfaces…
no need to crop the mirror part, actually even the non-mirror part inverted can be aligned (once inverted it becomes the mirror-space). I know it sounds a bit weird, but, actually it works. The only real issue is the collisions between spaces… that really cut holes.
I got it.
I’ll save the results and send GDrive link
This is a difficult room to be sure. Everything is against you. Everything is glossy/reflective. Lots of repeating patterns. Light is less than ideal. And its cramped. Basically everything that your told to avoid is in this room.
A couple of tips for better results next time.
Use ISO 100. A lot of the bumpiness in the model is from the camera noise.
Don’t use aperture priority mode. The shutter speed is changing between shots. In one shot the toilet or sink is bright and then higher up with the window or mirror light the camera’s exposure adjusts to the brighter light and the toilet/sink are now very dark. If you was shooting manual the toilet and sink would be the same brightness. Reality Capture is dumb and will think these 2 things are completely different subjects.
White balance is constantly changing. Outside is very blue so it makes the wall by the toilet very yellow to compensate. The light on the ceiling is very pink so auto white balance adds more green, and the light by the mirror is green and again auto white balance is compensating. Each frame looks good by itself but RC is dumb and can’t tell it is the same object if in one frame it is blue green and the next red orange. If you shoot raw you can adjust to one white balance that looks good across the most pictures.
Overlap. It looks like mostly a grid like pass on each wall shooting perpendicular, which is fine. Your scanning method looks good. One place i had a hard time though is finding common control points from one wall to the next. Example: I could only find 3 pictures of the corner that the wall with the door and the wall with the window meet at the top and none at the bottom corner. For this room what would have helped is the diagonal method standing in one corner shooting the opposite. The each wall by its self was covered good, but where they meet was hard for me to find.
I’m sure, with the raw files, I could produce better results.
Holy crap! Thanks so much!
I owe you one here.
Thanks for all the tips as well. You are right, I was fighting between manual mode and aperture mode. And one thing I learned after doing this location was to just use a fixed white balance for the exact reason you mentioned
I think another tip for myself or anyone who is in a tight space like this… is to pick a corner of the room and just start shooting across diagonally and make the movement up (Go up a foot every picture so you can capture more of the room)
Which settings were you using to align? Here is my goto settings right now. (Only thing I change is the overlap if needed) And I started using the reprojection error at 1… I think this makes for better texturing later on but Im still playing with it.
OK, the shooting system I was told for room interiors is:
For each shot that we’re used to doing, make it three shots i.e. pointing middle left and right maybe 15o each way. That makes a panorama of three, which we well understand won’t support ea other because no points are seen from two or more different camera positions; the position is same for all three.
But having taken panorama of three no.1 (1A,1B,1C), side-step - he says by one third (the 3x3 grid in the viewfinder) not by 20% (i.e. the RC recommendation of 80% overlap) and take another panorama of three no.2 (2A,2B,2C). And so on.
Then (this is my thought) 1C, 2B and 3A support ea other (they cover approx same picture area but from three different camera positions), as do 2C, 3B, 4A and 3C, 4B, 5A and so on. Draw it out on paper and it’s fairly awesome.
Then do it again with one third vertical overlap - but still only pointing 15o left and right - not suggesting 15o up and down as well, tho that wd be logical, and maybe a framework for taking ‘lots of extra photos’ if found necessary.
Haven’t thought it through yet but seems to me this gives a lot more views of each point than the 80% overlap recommendation - but crucially, each point seen from a wider range of camera positions, while still close enough (15o on view centre line) to the ideal ‘perp to wall face’ (esp if the wall face isn’t all that orderly). But it’s ‘lots more views’ in a way that can be snapped quickly, with 3x as much quick panorama-swinging and just over half as much slower side-stepping, esp with tripod.
A separate suggestion I was told, at internal corners of the room, take lots of transition shots between the two planes - a series in 15o steps, which sounds like six but can perhaps be just four extra to endmost 15o-sideways shots of the ‘panorama of three’ pattern previously described.
Not saying this would help with this thread’s repeating-pattern problem - more that hopefully it would get nearly all photos to Align together sooner in one Component.
Tom, yeah your right. The corner overlap was lacking and this would have helped. I also use this pano method to gather the same amount of “information” but without having to move my tripod as much.
You did that already - despite knowing that panoramas are supposed to be a no-no? That means you realised it would work something like I described?
What angle to left and right did you use - like the above 15o?
What sidestep overlap (tripod moves) did you use - RC’s 80% or the above 67%?
Tom,
Yeah, look at the my first post on this thread. The not the drawings but the screen shot and you can see the pano or partial panos in some cases. You have to think this software is just doing, at its simplest and at the core, is just doing triangulation billions of times over.
In my beautiful art work below photo 1 and 2 would easily align but the accuracy of the black dots position would not be that great. On the other extreme photo 1 and 3 would not easily align as the difference in similarity is too great and the software would take too long and run out of memory trying every possible combination. but if they did align ( maybe with the help of control points ) the accuracy would be very great
Now instead of taking a bunch of intermediate shots and moving a tripod to many positions, I could achieve high accuracy with overlap pano pictures and minimal tripod movement.
Again this is way over simplified and I wouldn’t use this method by itself but it does lower the total pictures i have to take and yields good results.
This is a v radical interpretation of the principle - the usefulness hidden in those shunned panoramas.
No question here of stepping along the wall, with close overlap. Strangely like the few, ‘random’ stations of laser point cloud scanning. I can see how it wd ‘lower total pictures’.
Just wait for the guys to log in on Monday - this one will run and run!