A 'good' photo set

I answered a Q in https://support.capturingreality.com/hc/en-us/community/posts/360008683312 but then thought it sholud be a new thread. Comment and better info would be welcome.

AFAIK - my interim discoveries in finding out what makes a ‘good’ photo set -

Sharpness - if using Depth of Field (rather than autofocus), set a higher standard - in https://www.pointsinfocus.com/tools/depth-of-field-and-equivalent-lens-calculator >Advanced Options, set Circle of Confusion to d/2000 not default d/1500. For best DoF, use small aperture like f11 but not too small like f22 or difraction sets in.

Freedom from Noise - use the camera’s ‘native’ slow ISO, usually ISO 100.

Those two together mean slow shutter speed - so use tripod if necessary, and flash for interiors.

Not sure whether blurred and/or noisy shots should be weeded out, or left in - ‘the more the merrier’.

Not sure whether improving the camera mpx - e.g. in my case from old Nikon’s 10.2mpx to my desired Fuji’s 24mpx - would help with the surface bubbliness. Of course, taking sq.rt of (24/10.2)x area improvement gives only a 1.53x linear improvement - not so dramatic.

Overlap/coverage - 80% overlap is recommended (assuming a single row of images) - i.e. move along one fifth of frame width ea time, meaning each feature in the view appears in five photos. If ‘tiling’ shots (i.e. a grid of shots overlapping both vertically and horizontally), move along and up/down one third of frame height and width, meaning each feature appears in 3x3 = 9 photos. The 3x3 composition guide grid in many viewfinders shd be useful - mine doesn’t have it unfortunately. Moving half frame, meaning each feature appears in 2x2 = 4 photos, might be enough.

RC relies on seeing any given feature from at least two (or 5 or 9 or 4) different angles. So don’t do panoramas (fixed camera position, panning around) because then any given feature, though appearing in several shots, is still seen from same angle. Move the camera for every single shot.

Those different angles must link together in a chain having not more than 30o difference, preferably less.

But preferably not less than 10o, def not less than 5o. RC must not see any pair of photos having angles of less than 5o difference, otherwise exponential errors arise in placing the tie point depthwise - basic trigonometry. So more photos is not better if they’re crowded together at less than 5o intervals - weed them out. In this case, not ‘the more the merrier’.

With a moderate wide-angle lens giving 67o angle of view horizontally x 47o vertically, the 80% horizontal overlap rule gives 13.4o angle difference if the camera is landscape, 9.4o if portrait. The one-third horizontal and vertical overlap rule gives 22.3o difference in the landscape axis, 15.7o in the portrait axis - ideal.

Both close-up and further-out shots are good but must be linked together by intermediate-range shots. Still experimenting with the necessary range-to-subject multiplier, and whether the linking shots must cover the whole scene, or just every-so-often.

Comment on the above scheme would be welcome.

“But preferably not less than 10o, def not less than 5o. RC must not see any pair of photos having angles of less than 5o difference, otherwise exponential errors arise in placing the tie point depthwise - basic trigonometry. So more photos is not better if they’re crowded together at less than 5o intervals - weed them out. In this case, not ‘the more the merrier’.”

Huge. HUGE. Flippin mahoosive piece of information…this means my practice of walking around the base of the tower I am modelling with my Pentax on 5 shots per second (or whatever it can do, I forget)…is NOT helping. I did same with GoPro on a pole to get different angle (4m or so, looking down) but that only does 1s/second so probably not so bad…but I get what you mean…I’ve given RC two images of the same point, from the same angle, which means it’s evaluation of that point’s depth could vary wildly…two good guesses are better than one…but two bad guesses are worse than one…and that confuses everything else.

Lightbulb moment. Each photo needs a noticeable difference of composition…if you have two that are “too similar”; delete one.

 

Thank you! If you need me, I’ll be going through 2,159 photos going “yep. Yep. Yep. Nope. Yep. Yep. Nope. Yep…etc”…

Caution! I’m a newby - this is how it looks to me but I’m looking forward to the guys’ feedback on Monday. Out of all the things I said above, this particular point may be quite wrong!

It’s not just that you’ve “given RC two images of the same point, from the same angle, which means it’s evaluation of that point’s depth could vary wildly” - it’s that triangulation off such a small base line (distance between two close-angle shots) sends a pair of barely-converging lines off into the distance, and it’s almost anyone’s guess how far away they’ll cross.

The potential error is in proportion to 1/tan or 1/sin of the convergence angle - at 30o 1.7, at 20o 2.8, at 15o 3.7, at 10o 5.7, at 5o 11.4 - and so on exponentially.

I’m a newby too…but what you say makes sense…given two near identical 2d images, the system is likely  nit much better at estimating the third dimension than you or I…(or eye, lol)…if in one image it’s a foot out, in the second it could be a foot out the other way…that’s doubling the error…so by taking out one image, you halve it…

But if the second image gives a better correlation…more images helps. It’s all about signal to noise ratio…or data to error, at least…the bigger your potential errors, the less more of the same data helps (assuming the chance of error is the same). What you need is different data (ie, different angle)

Related question. All the images I use are jpgs. My Pentax can shoot RAWs…any advantage to using them? CAN RC use them?

The question is - if a particular type of measurement is inherently liable to errors, is it best to eliminate all such instances (when better types of measurement are readily available), or to allow lots of them, in the hope that their errors will balance out? The latter may be false, especially if there’s some form of bias in play.

Dear Nigel,

RC can process RAW images. Please check allowed input formats https://support.capturingreality.com/hc/en-us/articles/115001527832-RC-s-input-import-file-formats

RAW imagery usually has advantage that you can adjust image parameters before processing in RC without loosing image quality. 

Regarding optimal image capture workflow, I’d recommend reading https://support.capturingreality.com/hc/en-us/articles/115001528211-Taking-pictures-for-photogrammetry

That link has much good advice but unfortunately doesn’t help when it comes to the special problem of not moving around the subject looking inward, but being inside the subject (a domestic-size room) and looking outward.

Especially when a lot of what’s seen is flat plaster (featureless in 2D) and smooth extruded shapes (featureless in 1D). I get round that by pinning little ‘rosebuds’ (scrumples of patterned wrapping paper) to flat walls and extruded door frames, which allows photos to quite easily register together as single Component (for my purpose it doesn’t matter that the ‘rosebuds’ appear in the textured model).

But even that doesn’t work to get me through open doorways so that two rooms simply register as a single Component https://support.capturingreality.com/hc/en-us/community/posts/360009260331 .

What about a RC official answer to the ‘no photo pairs less than 5o apart’ suggestion above?

Regarding RAW images:

Windows nowadays comes with built in support for a wide variety of raw image formats. If windows can open and display the raw format image, then RC can very likely load and use it.

Regarding taking images inside a room:
Walk around the room with the wall to your back and take pictures of the opposite wall. Don’t stand in the middle and rotate around.

The latter is suitable for panorama shots but not 3d reconstruction.

“Walk around the room with the wall to your back and take pictures of the opposite wall. Don’t stand in the middle and rotate around.” True, but too elementary (no offence), other than in an empty, large room.

In a domestic-size inhabited room with lots of clutter occluding views of the ‘far’ wall, a great deal of adaptation of that ideal advice is needed, which needs v clear discovery and application of principles (i.e. what RC wants to see) rather than rule of thumb.

Nigel - I think I must withdraw my “RC must not see any pair of photos having angles of less than 5o difference” advice.

Though I picked it up from somewhere, applied it, and it seemed to help, I’ve since noticed too much advice to contrary e.g. see main pics in https://support.capturingreality.com/hc/en-us/articles/115001569011-Working-with-Components-Merging-components .