Forum

Author Topic: Processing a massive underwater photoset without GPS  (Read 5263 times)

Mariusz_M

  • Newbie
  • *
  • Posts: 34
    • View Profile
Processing a massive underwater photoset without GPS
« on: October 23, 2021, 09:57:39 AM »
Hi,

I am processing a massive underwater photoset (so far 50 000 photos) and I keep taking 10 000 photos every day. There is no GPS data and I am looking for a solution to speed up the matching part of the process. Therefore, I need to understand how each of the following work.

1. Reference preselection set to sequential + generic preselection. What is it going to do? Will the program do the sequential preselection first and only the photos that have no pair will have generic preselection applied? Or will it do the generic preselection for all photos, finds all matches and then also add sequential preselection on top to find even more matches and actually extend the length of the process?

2. Reference preselection set to seqential while every second photo disabled, so only half of the photos are being aligned. Provided that they will align somehow... Then I enable the other half of the photos and what are my options then? Sequential again? Or perhaps Estimated? With generic preselection on or off?

Any other ideas how to match only the photos that should be matched?

CheeseAndJamSandwich

  • Full Member
  • ***
  • Posts: 153
    • View Profile
    • Sketchfab Models
Re: Processing a massive underwater photoset without GPS
« Reply #1 on: October 23, 2021, 12:25:11 PM »
I'm in a similar (dive) boat to you...
I'm up to 24,000 photos with one of my sites.
I'm scanning whole dive sites with pair of GoPros on a pole (2x quicker scanning).  And obviously, no georeferencing data at all.  Plus i only have Metashape Standard.

What camera(s) are you using?
Swimming/manual or ROV/AUV?
How or you taking them? Time-lapse?
What's the site like?  Rocky? Reef? Wreck?
What are the conditions like?
What's the aim of the scanning?  Accuracy needed?

I started off with very random scan tracks from each dive... But have learnt to do a lot more 'mowing the lawn' style scans.  Much quicker to cover the area, but also much quicker for MS to align.
As my early scans are all over the show, they overlap a LOT!  Meaning my Align Photos processes were taking several days to process...  The mesh, only half a day.
So definitely optimise the scanning such that you have good coverage, with the correct amount of overlap, and don't go over too much you've scanned before...

The results i get from the individual scans are pretty amazing straight out of Metashape... Don't need to do much to them.
As I'm scanning with 2x GoPros on a 3m pole, swimming 4-5m above the reef/rocks, shooting time-lapse photos with a 2 second interval (3 would be nicer, but can't set it)... I get a really good quality scan, especially the are between the two cameras...  MS seems to get all the lens calibration info out of the exif tags, such that i haven't seen any distortion of the meshes/models, no bowing, dishing, anything! They've all been amazingly accurate, even over 0.5 km drift scans.
But obviously presented in some weird orientations and scaling (i've requested MS Standard gets a basic reference system such that we could put down 3 control points.)

What I'm doing to combine all the 'missions', dives, scans, is to create a model of that scan, with align set to Low, 20k, 2k... Then do a bit of Gradual Selection, only to 15 tho... Then create a mesh from the sparse point cloud, which takes seconds, and then run Reduce Overlap, set to 9... Remove 'Disabled', then optimise cameras. This removes about 1/3 of my cameras, which is inline with 2s time-lapse interval being too fast for my swimming, 3s probably being better...  Then create the mesh from Depth Maps to see how it look.  Which usually is perfect. So i land up with a good scan, with a thinned out set of cameras, leaving just one right amount for the scan.

I then bring all the individual scan photo sets into one project, putting them in camera folders...  Then i run a big alignment the better settings, High, 40k, 4k...  this takes 2.5 days still... but it's a lot quicker than it was gonna be!
Once that's finished, i then repeat the steps of optimising the point cloud and reduce overlap again...  Then the mesh/model only takes 12 hour to calculate...  Pretty Quick!  And the resulting model comes out great.

I've downgraded to 1.7.3, as suggested by other forum users, as 1.7.4 & 5 left big holes and lots of blobs where there was weak data (shaded side of rocks, for me).

The next stage i'm processing, trying right now, is incrementally adding 1 more scan to an already aligned model of 9 scans...  You have to have Keep Key Points turned on in Preferences!!  Before you start!!
The alignment has just finished, and the point cloud looks good... will set it processing the mesh in a few minutes, to see what i get...

This incrementally adding extra scans I'm gonna be weary of, as the optimising steps will reduce the points, cameras in the existing areas away from the new data, each time i add another scan...  So what i think i might have to do is occasionally do a full alignment of all the, say 15, sets of scans, from scratch...  (A bit like a computer backup regime....  Full Backup, Incremental, Incremental, Incremental, ... Full Backup, Incremental...  etc...)  I'll still be only combining the optimised set of data for each scan....  but not degrading the whole model each time i add a bit more.... I'll have to suffer the 3 day alignment processing occasionally, but it'll be fine...

Bzuco

  • Full Member
  • ***
  • Posts: 178
    • View Profile
Re: Processing a massive underwater photoset without GPS
« Reply #2 on: October 23, 2021, 03:36:00 PM »
@Mariusz_M
Generic preselection will do additional quick calculation after detecting points phase.
It is usefull for next phase - matching points between photos, as it will discard photo pairs from process which don't have any overlap/common features. Without gener. presel. matching points would be between each photo with each other( num photos * num photos tasks).
As for the "estimated" and "sequential" options, I am not 100% sure, from my perspective is just let metashape know how we were taking photos, if in chaotical order(orbiting some object), or in sequential(e.g. long wall from left to right). I could be wrong and it is more behind this options.

@CheeseAndJamSandwich
Your underwater work is interesting.
Are you using generic preselection? and what is your hardware? 2.5days for one big alignment seems to be too long. I would expect several hours of alignment time for ~25000 big photoset.
Did you try decrease key point limit to e.g. just 20k?
What precision you need in the end?(pixels on texture per cm/m...or pointcloud points per cm/m).
Are you many times also orbiting aroung rocks when taking photos, or is everything mostly from top down?
And last funny question: can you swim faster to better use 2s interval on gopro? :D

Mariusz_M

  • Newbie
  • *
  • Posts: 34
    • View Profile
Re: Processing a massive underwater photoset without GPS
« Reply #3 on: October 23, 2021, 04:26:33 PM »
Thank you guys for your lengthy responses.

I have been doind underwater photogrammetry for several years now and processing 20 000 - 30 000 photos in one chunk is not new to me. I know how generic preselection works and how to scan my underwater sites. I already have my workflows for my very precise underwater work and I really know what I am doing with almost all the options I use... Apart from the Reference preselection SEQUENTIAL and ESTIMATED that are not really described in detail in the manual. Furthermore, the way these work have changed over the last few years, so I really need information from someone who is sure.

The difference between this project and my previous big projects is that I will probably end up with around 120 000 photos and I want to find an optimal way before I just click "Generic Preselection" and "Reference preselection" and then wait for a week to see what happens. Also, gathering photos will take a while and I need to be able to quickly do some preview alignments to double check the coverage.

So my questions are very specific and technical and only about the alignment parameters. Let me rephrase them then:

Does anyone know how Reference Preselection Sequential and Estimated work without GPS and how they work in combination with Generic preselection?
Which one is applied first?

If Generic is as second, then does it take into consideration pairs already found by Reference preselection or does it do its thing to all photos anyway?
If does take into consideration the pairs found by reference preselection, what happens then? Does it only use generic for photos without pairs?

I really need to understand the algorytms behind each of these options and their combination. Anyone?

Bzuco

  • Full Member
  • ***
  • Posts: 178
    • View Profile
Re: Processing a massive underwater photoset without GPS
« Reply #4 on: October 23, 2021, 06:12:30 PM »
OK, understand, sorry for slightly hijacking this thread.

I just made quick test on 200 photo set(house and backyard from ground) with combination of params and resulted time.
"Not all aligned" - model looks good, but not all cameras were marked as aligned.
So +1 to know, when and why use exact methods and for what scenario.

CheeseAndJamSandwich

  • Full Member
  • ***
  • Posts: 153
    • View Profile
    • Sketchfab Models
Re: Processing a massive underwater photoset without GPS
« Reply #5 on: October 23, 2021, 09:20:24 PM »
@CheeseAndJamSandwich
Your underwater work is interesting.
Are you using generic preselection? and what is your hardware? 2.5days for one big alignment seems to be too long. I would expect several hours of alignment time for ~25000 big photoset.
Yes.  It's on by default, tho i did do some tests a long time ago and it was hilariously longer without it.
I have a big fat ThinkPad P51, from 2018...  Sadly no desktop...
https://www.notebookcheck.net/Lenovo-ThinkPad-P51-Xeon-4K-Workstation-Review.229212.0.html
I have bags of RAM in it, 64GB, but my jobs hardly use more than 20, perhaps 30GB.
But the old Xeon and the crappy old Quadro M2200 in it just don't compare with todays desktop HW...  I continue to dream of course....

As my scans started off pretty erratic in their paths for the early scans, i have a lot of overlapping, areas scanned more than once...  This turns out to make the alignment times go exponential.
My next test plan is to compare each of the individual scans, and delete the photos that cover the same areas, keeping the ones from a better scan...  Perhaps when the water visibility was better...  And deffo on the areas where the rocks/reef meets the sand... As you can go back the next day, and if the swell was big, the height of the sand might have risen, dropped a foot!  Which causes horrible double layers in my mesh.  With a bit of thinning out of the duplicated areas, i could probably knock a day off of the big alignment job.

Quote
Did you try decrease key point limit to e.g. just 20k?
Yes...  BUT... I was getting holes in my model where the the data was week (shaded sides of rocks, etc)...  But after hearing about 1.7.4 & 5's bad meshes, and downgrading, rerunning the mesh and the holes filling in nicely, i think it was that, rather than the reduced point limits...  My current job is adding a 10th scan to an already aligned 9 scans, that was done with 40k, so this is 40k too...  The alignment was obviously a lot, lot quicker, and in a few hours time I'll have the mesh/model to check...
I'll test 20k for a future big alignment.

Quote
What precision you need in the end?(pixels on texture per cm/m...or point cloud points per cm/m).
I don't need great accuracy at all...  It's for printed off 300dpi maps on A4, A3, and perhaps some massive posters for the dive shops...  Also video flight around the models, Sketchfab models (just using the free 100mb limit, which is actually enough, deffo enough for loading onto a phone!)
I'm not gonna be doing any high precision measurements or analysis from this data.  But it has to look correct of course.
Luckily, using my two GoPros, the models have come out perfect, no dishing, no bananas, even over long drift scans, that might be 500m long...  It all just comes out great.  Getting these good results without anything special, much work is actually very, very convenient.

Part of this work is also supporting the coral reef rehabilitation efforts we have going on here too...  So mapping out bubble sites and then monitoring growth of the coral frames we put down.  So i might be doing more smaller scale scans, of 2m x 2m frames, of individual corals, etc...  Then it could become increasingly important to get cm, mm accuracy, but i'd need the professional version to allow any dimensional analysis of these.
I could deffo do with some of the other Professional features like DEMs, etc, which naturally work great as dive site maps, but simply will never ever be able to afford the $3,500  :'(

Quote
Are you many times also orbiting around rocks when taking photos, or is everything mostly from top down?
Some of the site dictates the orbiting around the rocks, as they're literally house sized, 10m tall, wide...  And in my earlier scans, i was zooming around all over the show, with little planning... But now i'm trying to do a lot more 'mowing of the lawn' scans, which are a lot, lot quicker at covering area... Though keeping straight UW is tough, even with a compass, and if the pass is long, then there's a chance track diverges from the previous pass' track, so you get an annoying scar of weak data... doh...  Underwater, everything is more difficult.
If i was to redo the whole job of the big site I'm doing right now, i could probably do the same area in 6 dives instead of 10.  Which would take a lot, lot less time to process, as there would be a lot less overlapping scans, which disastrously affect the alignment time.

I have 3 or 4 other sites i need to scan soon, and a couple of shipwrecks i have scans of...  It's an awesome project to be doing, and learing MS has been a lot of fun... If only i had a 5950X and a 3090!!!

Quote
And last funny question: can you swim faster to better use 2s interval on GoPro? :D
LOL  I'm swimming as fast as i can!!!  :P
I'm using my big long freediving fins/planks, but also have a sling tank, a whole extra AL80 tank on my side, as I'm scanning alone, so have to double up on everything for safety...  So not the most streamlined underwater... Can't find a buddy that can keep up with me! (or that also has two cameras).  Oh, and I'm chugging down both 200 bar, 11 ltr tanks, down to 30 bar, in 45 mins!  My extra layer, kilos of 'bio-prene' might not be helping either... Though it keeps me warm.
The limited choice of interval is a real bummer actually...  3s would just make everything so much easier!  I've requested it on GoPros forums, Basically just choose any value of seconds... but that's just a dream.
« Last Edit: October 24, 2021, 11:40:50 AM by CheeseAndJamSandwich »

CheeseAndJamSandwich

  • Full Member
  • ***
  • Posts: 153
    • View Profile
    • Sketchfab Models
Re: Processing a massive underwater photoset without GPS
« Reply #6 on: October 23, 2021, 09:31:51 PM »
I have been doind underwater photogrammetry for several years now and processing 20 000 - 30 000 photos in one chunk is not new to me. I know how generic preselection works and how to scan my underwater sites. I already have my workflows for my very precise underwater work and I really know what I am doing with almost all the options I use...
If you can share any of your learnt wisdom, that woule be amazing.


So returning to your OP...
Quote
So my questions are very specific and technical and only about the alignment parameters. Let me rephrase them then:

Does anyone know how Reference Preselection Sequential and Estimated work without GPS and how they work in combination with Generic preselection?
Which one is applied first?

If Generic is as second, then does it take into consideration pairs already found by Reference preselection or does it do its thing to all photos anyway?
If does take into consideration the pairs found by reference preselection, what happens then? Does it only use generic for photos without pairs?

I really need to understand the algorytms behind each of these options and their combination. Anyone?
+1
I've been wondering what these do myself.  Though my trials didn't show changes in speed or results...  How does sequential know the sequence... Do the filenames have to be numbered sequentially, or just alphanumerically ordered?  And do my gopros contain any stuff in their exif tags that help MS, as when we're underwater, the GPS they contain won't work, and is normally turned off.

As i've addressed a few times in my other recent posts... There's a lot of features that simply aren't explained, not even in the manual...  Having a good explanation with some examples of uses, benefits, effects for some different styles of photogrammetry, would be amazing... And all us metashape users would benefit, metashape would be even easier to use, and it'd sell more copies for sure.

CheeseAndJamSandwich

  • Full Member
  • ***
  • Posts: 153
    • View Profile
    • Sketchfab Models
Re: Processing a massive underwater photoset without GPS
« Reply #7 on: October 24, 2021, 11:51:07 AM »
The next stage i'm processing, trying right now, is incrementally adding 1 more scan to an already aligned model of 9 scans...  You have to have Keep Key Points turned on in Preferences!!  Before you start!!
The alignment has just finished, and the point cloud looks good... will set it processing the mesh in a few minutes, to see what i get...

This incrementally adding extra scans I'm gonna be weary of, as the optimising steps will reduce the points, cameras in the existing areas away from the new data, each time i add another scan...  So what i think i might have to do is occasionally do a full alignment of all the, say 15, sets of scans, from scratch...  (A bit like a computer backup regime....  Full Backup, Incremental, Incremental, Incremental, ... Full Backup, Incremental...  etc...)  I'll still be only combining the optimised set of data for each scan....  but not degrading the whole model each time i add a bit more.... I'll have to suffer the 3 day alignment processing occasionally, but it'll be fine...
Well, the results of adding the 10th scan to the already aligned 9 scans didn't turn out that well...
There's a fair bit of double-skins over the rocky areas, where the old scan data and new scan data overlap... usually this comes out ok, unless it's sand, that has moved... But the rocks haven't.
So i'm gonna try aligning all 10 optimised scan sets from scratch, but using low/20k/2k, and hoping the bad results i'd gotten the last time i tried this was 1.7.4/5's dodgy meshing...
Hoping the low/20k/2k will shorten the projected 80 hours down to something more bearable...


Meanwhile...
Back to us wondering what reference preselection set to sequential or generic preselection do, and if it's relevant to us scanning underwater, without GPS data.

Kiesel

  • Sr. Member
  • ****
  • Posts: 332
    • View Profile
Re: Processing a massive underwater photoset without GPS
« Reply #8 on: October 25, 2021, 10:47:17 AM »
Hi,

from the manual page 30:

Quote
Generic preselection

The alignment process of large photo sets can take a long time. A significant portion of this time
period is spent on matching of detected features across the photos. Image pair preselection option may
speed up this process due to selection of a subset of image pairs to be matched.
In the Generic preselection mode the overlapping pairs of photos are selected by matching photos
using lower accuracy setting first.

Note

Using only Reference preselection mode without Generic preselection fits best the case when
only a few tie points are detected during the image matching stage, it may happen, for example,
for aerial surveys of forested areas or crop fields.

Reference preselection

In the Source preselection mode the overlapping pairs of photos are selected based on the measured
camera locations (if present). For oblique imagery it is necessary to set Capture distance value
(average ground height in the same coordinate system which is set for camera coordinates data) in the
Settings dialog of the Reference pane to make the preselection procedure work efficiently. Capture
distance information must be accompanied with yaw, pitch, roll/omega, phi, kappa data for cameras.
Rotation parameters should be input in the Reference pane. Then the preselection based on the new 3D
points calculated as original 3D point and vector in the direction of the camera view with the length
equal to the input Capture Distance value.

The Estimated preselection mode takes into account the calculated exterior orientation parameters
for the aligned cameras. That is, if the alignment operation has been already completed for the project,
the estimated camera locations will be considered when the Align Photos procedure is run again with
the Estimated preselection selected.

When using Sequential preselection mode the correspondence between the images is determined
according to the sequence of photos (the sequence number of the image) it is worth noting that with
this adjustment, the first with the last images in the sequence will also be compared.

Regards

CheeseAndJamSandwich

  • Full Member
  • ***
  • Posts: 153
    • View Profile
    • Sketchfab Models
Re: Processing a massive underwater photoset without GPS
« Reply #9 on: October 26, 2021, 02:02:30 PM »
Hi,

from the manual page 30:

...

Regards
We've all RTFM.
But we need to know more details about these features, and also what would be considered the best settings to use in our situation...  Namely, aligning thousands of modes taken underwater, where the camera's photos contains no GPS/reference data...
As our alignments take 3 days to a week, we really, really would like to know what will help us...  As the savings could be measured in days.  That's potentially a lot of time, electricity and opportunity cost.

SimonBrown

  • Jr. Member
  • **
  • Posts: 62
    • View Profile
    • AccuPixel Ltd - Dealer and Training Centre
Re: Processing a massive underwater photoset without GPS
« Reply #10 on: October 28, 2021, 06:53:10 PM »
Unless the GPS data is accurate then adding it won't help.

Test this by shooting a dataset with an iPhone...then use Reference Preselection set to source and watch what happens.

Last time I tested it, a mess as the embedded GPS values were good enough to tell me roughly in the world where the image was shot...and inaccurate enough to confuse Metashape.

We do work with underwater GPS and merge the camera positions into the JPEG data...but its never used for preselection. We have gone as far as writing a tool to merge GPS data with images - thats how useful it can be - but not for preselection. See https://accupixel.co.uk/2021/07/26/new-release-gps-position-processing/ for more details.

Sequential might work if you are very disciplined with the sequence of image creation, but it will not match adjacent images in parallel runs. Again, for good reasons we don't use this underwater.

So what to do?

First step is to look at the process of gathering images - are you shooting stills? Or deriving still images from video?

This is a fundamental question. Video is popular underwater, but it carries a whole set of issues still images simply do not generate.

So its worth understanding that before going further - what method do you use?
Agisoft endorsed online Metashape training - see: https://accupixel.co.uk/metashape-training-courses

CheeseAndJamSandwich

  • Full Member
  • ***
  • Posts: 153
    • View Profile
    • Sketchfab Models
Re: Processing a massive underwater photoset without GPS
« Reply #11 on: October 28, 2021, 08:13:24 PM »
Sequential might work if you are very disciplined with the sequence of image creation, but it will not match adjacent images in parallel runs. Again, for good reasons we don't use this underwater.
Do you not 'mow the lawn' at all?  Or not have multiple cameras on pole/frame?  Or not use sequential for UW?

I've wondered this, and i think i tested it, separating each camera into it's own camera folder, but I'm not sure it made any difference in my test...
Do the photos have to be named sequentially? img00001, img00002, etc... Or does MS just load them alphanumerically, and just take that as the 'sequence'?

Sequential would not like my mixing of tracks I guess...  as it'll be alternating from left to right to left...
Perhaps i need to do some more tests with sequential file names in separate folders...

My results otherwise, ignoring processing times, are actually very good from having the two parallel tracks from the two cameras on the pole.  The data between them is very good... I'd consider adding a 3rd in the middle if i was doing some closer, high resolution scans...  You do land up with 3 perfectly parallel, and perfectly overlapping runs, better than 'mowing' the lawn could achieve.
Which reminds me, i must calibrate my 3m pole/camera pitch vs. 'altitude' vs. swimming/drifting speed...  Workout the resulting overlaps...  I know the speed give me about a third too many photos with this setup, and that date between the cameras is excellent...  So it's good enough... 

Quote
So what to do?

First step is to look at the process of gathering images - are you shooting stills? Or deriving still images from video?

This is a fundamental question. Video is popular underwater, but it carries a whole set of issues still images simply do not generate.

So its worth understanding that before going further - what method do you use?
Time-lapse Photo mode in the GoPro 7&8, so 12MP photos every 2 seconds.  Looking down or normal to the reef/rock, from about 4-5 above the bottom, or away from the rock.
Never tried the video source method...  can't afford the disk space needed to store 45 mins of 4k footage!!!  lol.  I guess it's great if a video is all you've got.
All other setting are automatic, and they do a good job with the UW WB these days... The 7 is solid, but the 8 i have does have an 8m-ish bug, where it gets confused and drifts back to topside WB, the photos turning blue, then back to 'good' UW WB again... annoying af.

So...
Alexey...
  • For Sequential, do the filenames matter??? Or folders?
  • When is using Estimated for a 2nd run of Align Photos worth doing?
  • Are their any other tips for Aligning these UW photos with no GPS data?
Thanks!

SimonBrown

  • Jr. Member
  • **
  • Posts: 62
    • View Profile
    • AccuPixel Ltd - Dealer and Training Centre
Re: Processing a massive underwater photoset without GPS
« Reply #12 on: October 29, 2021, 11:43:02 AM »
Quote
Sequential would not like my mixing of tracks I guess

Mine neither. But that is not the reason I do not use it.

To reiterate
Quote
...it will not match adjacent images in parallel runs
is the reason.

Sequential works well with planned mission drone flights. We divers are not drones. Its not going to fix your issue, but if you want to try it then I believe its based on the EXIF image creation time, not filename or folder structure.

Each camera image should align to its neighbour. But no alignment will occur between the camera image sets using sequential. A second alignment will be needed with Reset Alignment ticked off to align all points.

And to reiterate:

Quote
Unless the GPS data is accurate then adding it won't help.

If you really want to go the accurate underwater GPS route then its possible. A budget minimum of perhaps $70~100k would see the issue fixed, but it will still come with limitations, setup and working methodology that would add time to the work on site.

Remember, everything can be fixed...but it all comes with cost.

If you have not used video then don''t bother - stills will deliver more data for less work.

You may be generating more images and therefore more data than is required to produce a result. The entire Thistlegorm - 7 acres - took 24,000 DSLR images (see https://deep3d.co.uk/2020/05/03/the-ss-thistlegorm-content/. More images is not always best...its the correct number we really need.

Shooting at 3 to 4m from the subject should provide massive coverage...thats far too distant for my style which is typically 1~2m to allow the strobe light to restore colour...my gut feeling is there are more images than really needed in the project.

There is no magic silver plated solution here.

There is a mass of data to work through, You could spend more on processing hardware, change the cameras...but the basic approach of how the images are captured might be the best place to start - and its zero cost beyond the time needed.
Agisoft endorsed online Metashape training - see: https://accupixel.co.uk/metashape-training-courses

James

  • Hero Member
  • *****
  • Posts: 748
    • View Profile
Re: Processing a massive underwater photoset without GPS
« Reply #13 on: October 30, 2021, 01:39:02 PM »
So...
Alexey...
  • For Sequential, do the filenames matter??? Or folders?
  • When is using Estimated for a 2nd run of Align Photos worth doing?
  • Are their any other tips for Aligning these UW photos with no GPS data?
Thanks!

I'd also love to know more about sequential preselection.

My guess would have been that choosing sequential additionally ensures that sequential pairs are always matched/compared, when they may have not been picked up by generic preselection, rather than excluding any non-sequential pairs.

That doesn't make complete sense though, because I would think that 99% of all photos taken by everyone with any hint of method to their madness would take one photo after another in an overlapping sequence, in which case it would make sense for metashape to do this regardless. And as the number of sequential pairs only increases linearly with the number of photos, there's hardly any processing time penalty for doing it.

Also my guess doesn't make sense because if it was additional, then you should be able to choose this in addition to estimated and source reference preselection types, rather than only being able to choose one of three.

Anyway, my guesses are no use to anyone, including myself, so I'd love to know what it's doing really.

CheeseAndJamSandwich

  • Full Member
  • ***
  • Posts: 153
    • View Profile
    • Sketchfab Models
Re: Processing a massive underwater photoset without GPS
« Reply #14 on: November 06, 2021, 06:56:45 AM »
Sequential works well with planned mission drone flights. We divers are not drones. Its not going to fix your issue, but if you want to try it then I believe its based on the EXIF image creation time, not filename or folder structure.

Each camera image should align to its neighbour. But no alignment will occur between the camera image sets using sequential. A second alignment will be needed with Reset Alignment ticked off to align all points.
OK...
So i finally got around to testing this, and it's something that i somehow missed in testing when i started this project, but i actually enabled Sequential when aligning the scan from 1 dive, and blimey, it zoomed through the alignment!  :) 8)  And i think it also dealt with the trickier alignment of the shallow (5m) images, where the dancing sunrays really, really confuse MS (i need to scan those areas on cloudy days... Wet Season is coming, so that'll be easy!).  I'm currently running tests right now on one of the problematic scans, that usually has a lot of non-aligned images, to quantify the times and results.

As you say, when i tried to do the Alignment on my big model, containing 12 'missions' (dives) of scans, that otherwise overlap 9by too much, doh), then only one of the scans aligned, leaving the other 11 with zero aligned.  I tried rerunning the Align Photos task, with various setting, but it failed to align a single extra camera.  I had Reset Alignment UNTICKED...  What settings here would possibly allow the other cameras to align?

Quote
There is a mass of data to work through, You could spend more on processing hardware, change the cameras...but the basic approach of how the images are captured might be the best place to start - and its zero cost beyond the time needed.
I'm having to live on less than £10 a day at the moment, as I'm proper skint now.  So i'm stuck with the cameras and hardware i already had  :'(

« Last Edit: November 06, 2021, 07:56:37 AM by CheeseAndJamSandwich »