0.9.0 1584 + Toshiba Qosmio versus 0.9.0 1542 + Asus P31S

[Traceologue]

First of all, congratulations for such an amazing and efficient software!! It makes high quality photogrammetrical reconstruction available for a large range of purposes. I am currently testing its potential in use-wear study of archaeological tools.

I recently wanted to compare the difference of alignment with and without pair selection (Generic/Disable) in high Accuracy mode and I got strange heterogeneous results from Photoscan Standard Edition 0.9.0 1584 installed in a TOSHIBA QOSMIO X500 (Intel Core i7 720QM, 8.0GB, NVIDIA GeForce GTS 360M). In fact, it seems that starting a new session (File/New) is not sufficient for clearing the memory since the subsequent Alignments and Geometries are then aleatory. Only by closing Photoscan (File/Exit) after each work, I can constantly get reliable reconstructions.

With the same sets of images and the same parameters but with version 0.9.0 1542 installed in an Asus P31S (Intel Core i3 2350M, 8.0GB, NVIDIA GeForce 610M) it is not necessary to close the software for successfully starting a new session. I don’t know whether the bug is due to the hardware or to the last release of Photoscan? I spent half a day before understanding that my set of photo-macrographs was not faulty, despite a low depth of field at 1.5:1 magnification (worn tip of a Palaeolithic quartzite tool), but that something else was wrong.

I hope this may help not going too fast in conclusion when making comparisons ...

[Alexey Pasumansky]

Hello Traceologue,

Thank you for reporting.

But have you checked the latest release (build 1586)? Does it still has similar problem on the Toshiba laptop?

[Traceologue]

Hi Alexey,

Thank you for your fast reply! Here the report of a new series of tests with release 1586.

First alignment after installation of release 1586 (Align high disabled):
-   wrong

Second alignment after restarting the computer (Align high disabled):
-   wrong but different

Third alignment after restarting Photoscan (Align high disable):
-   wrong, but different from 1rd and 2nd

Fourth alignment after restarting Photoscan (Align high generic)
-   fine

Fifth alignment after restarting Photoscan (Align high generic)
-   fine

Sixth alignment after restarting Photoscan (Align high generic)
-   fine

Seventh alignment WITHOUT restarting Photoscan (File/New - Align high generic)
-   fine

Geometry of seventh alignment (Geometry sharp high Face count 0 Filter threshold 0.3)
-   GREAT

Eighth alignment WITHOUT restarting Photoscan (File/New - Align high DISABLED)
-   fine

Geometry of eighth alignment (Geometry sharp high Face count 0 Filter threshold 0.3)
-   GREAT

Ninth alignment WITHOUT restarting Photoscan (File/New - Align high DISABLED)
-   wrong but different from 1st, 2nd and 3rd

Tenth alignment after restarting Photoscan (Align high DISABLED)
-   fine

Geometry of tenth alignment (Geometry sharp high Face count 0 Filter threshold 0.3)
-   GREAT

   So far I cannot say what is faulty: the computer or the version of Photoscan. Is this possible to install two different versions of Photoscan in the same computer?

             Best,

            Hugues

[RalfH]

Hi Traceologue,

I have been experiencing the problem of Photoscan finding different solutions for exactly the same set of images since version 0.6 (when I started using Photoscan). It seems that the alignment is based on an algorithm with a stochastic component in it.

When I see that the images don't align as they should, I usually reset alignment for all images, select all images with the exception of those which I suspect cause problems (e.g. due to self-similarity, low viewing angle), start alignment for only the selected images. If the result looks fine, I try to align the remaining images one by one. This seems to force Photoscan to align the remaining images to the existing sparse point cloud. It usually works fine, and I just leave out any images which continue to cause problems.

For some reason, close-up images in general seem to be a bit tricky. This may be due to the very narrow depth of field. Perhaps try focus stacking (DOF stacking), e.g. using CHDK. I've been thinking about trying this but haven't found time to do it.

[Traceologue]

Hi Ralf,

Thanks for your comments and suggestions! I must acknowledge that my set of photos is difficult to process because of the narrow depth of field and I have been surprised to see that Photoscan is able to align them correctly, at least when running in my Asus P31S. With this computer the quality of the reconstruction is constant. Irregular results occur only when using the latest releases of Photoscan  in a Toshiba Qosmio X500.

The smaller is the depth of field the more efficient is focus stacking. However in the present case of the worn tip of a lithic pick focus stacking would request too many shots for giving its morphology.  Around real size (1cm real = 1 cm on the sensor) I feel that photogrammetry can produce models with better resolution.

[FoodMan]

if you stack photos, I can only recommend to stack ALL of them ... from a last project, mixing stacked and non-stacked photos on the same chunk has alignment problems...

f/

[RalfH]

Hi Tracelogue,

I did not mean to use focus stacking to get the morphology; I would expect that this does not provide a good depth resolution. What I meant was to use focus stacking to create images that are sharp irrespective of the distance from camera to object.

FoodMan, what is your experience with this so far? Is it worth the effort?

[Traceologue]

Hi RalfH,

   Sorry for not having well understood your suggestion. Yes, processing hyperfocused photos with Photoscan can gives a worthwhile result (here attached a detail of http://www.agisoft.ru/forum/index.php?action=dlattach;topic=533.0;attach=214 from http://www.agisoft.ru/forum/index.php?topic=533.msg2325#msg2325), but it is time consuming and you are quickly dealing with gigabytes of photos.

[FoodMan]

hey, yes sure focus stacking works.. but yes time consuming.. I use this app to stack.. not sure it's the best..

http://www.heliconsoft.com/

I was just saying that it's best to use ONLY focused stack images for a project.. I did one with both focused stacked and normal images, and the reconstruction had some artefacts;  mainly due apparently to misaligned images...

f/

[RalfH]

Great. Good to know that this works.