Worse sparse cloud with higher aligment setting.

[saproling]

Hello All,

I've been working with a model that produced a really good looking sparse cloud with the low setting on photo alignment, but the sparse cloud for the high setting was awful. I tried removing images that did not have GPS locations or if they had a high error values, but the sparse cloud came out even worse... Any suggestions?

Thanks,
S

[Porly]

Hello,

yes it is possible. For Aerial imagery I am using 40k keypoints per photo, for terrestial with poor light conditions I always gain better results with 20k points per photo.

Seems like in case of lower amount of points Photoscan use points or matches of higher precision (lower variance or higher interest value, better reliability). The lower amount of points causes more robust results in "poor" quality input data.

Cheers

Paul

[saproling]

Thanks for the tip! I'll definitely keep that in mind as I process more models.

-S

[chadfx]

Yes, cheers Paul, that just worked perfectly for some very quickly shot humanoids. ;-) I was getting much worse camera alignment with more points in the camera alignment.

Thanks! -Chad

[ozbigben]

Did you try gradual selection on the sparse point cloud?  I've had better results from using more points and weeding out the bad ones before generating the dense cloud.

[Marcel]

As far as I know the quality setting in the Align dialog sets the downscaling on the source photos before extracting points. When set to High, it uses the photos at original resolution, at Medium it downscales them to 50% size,  and at Low to 25% size.

The reason you get bad results at High is probably because your photos are not completely sharp, so Photoscan has trouble finding good points.
When you set the quality to Low, the source photos are downscaled to 25% of its original size, which means even a photo with a slight blur will be super sharp. So the downscaling improves the quality of your photos, leading to a more accurate Alignment.

While this may help, it's always better to have good source photos. Take a look at your photo at the pixel level (especially in the corners) to see if they are sharp enough. If they are blurry, moved or noisy then you need to work on your capture technique.

[Porly]

As far as I know the quality setting in the Align dialog sets the downscaling on the source photos before extracting points. When set to High, it uses the photos at original resolution, at Medium it downscales them to 50% size,  and at Low to 25% size.

The reason you get bad results at High is probably because your photos are not completely sharp, so Photoscan has trouble finding good points.
When you set the quality to Low, the source photos are downscaled to 25% of its original size, which means even a photo with a slight blur will be super sharp. So the downscaling improves the quality of your photos, leading to a more accurate Alignment.

While this may help, it's always better to have good source photos. Take a look at your photo at the pixel level (especially in the corners) to see if they are sharp enough. If they are blurry, moved or noisy then you need to work on your capture technique

Hello marcel,

downscaling the source photo is used for quality setting in the dense image process (build dense point cloud). In the alignment step, only the amount of interest points will be reduced, depending on the quality setting. High=40000 (with default setting) Low~4000. You can follow this in the console during "detecting points".

[James]

downscaling the source photo is used for quality setting in the dense image process (build dense point cloud). In the alignment step, only the amount of interest points will be reduced, depending on the quality setting. High=40000 (with default setting) Low~4000. You can follow this in the console during "detecting points".

I'm not sure about that. Just testing with some 36MP images, leaving point limit at 40k i get ~40k points per image using high quality, and ~5k at low quality, which agrees with you, but when i up the limit to 200k i get about 150k points per image on high, but still only 5k on low, which makes me think 5k is the limit on low because the images have been downsampled, and nothing to do with what point limit i set. I couldn't see anything in the console to prove it one way or the other however

[Porly]

You are right, James! Just tried to start align photos on point limit = 2000 @ high, what causes 2000 point per photo. Custom 2000 on low caused 2000 per photo too.

So looks like there is as fix min/max. value of interest points depending on the set point limit. But I dont believe that there fotos are downscaled... good, maybe I am wrong  

[Porly]

You are right, James! Just tried to start align photos on point limit = 2000 @ high, what causes 2000 point per photo. Point limit 2000 on low caused 2000 per photo too.

So looks like there is as fix min/max. value of interest points depending on the set point limit. But I dont believe that the fotos are downscaled... good, maybe I am wrong

[Porly]

Hello,

Accuracy
Higher accuracy setting helps to obtain more accurate camera position estimates. Lower accuracy
setting can be used to get the rough camera positions in a shorter period of time.While at High accuracy
setting the software works with original photos as are, Medium setting cause image reduction by factor
of 4, and at Low accuracy source files are reduced by 4 times more.

This is from the user manual, so marcel was right with his theory!

So that the input data is downscaled, and less number of keypoints are used, what causes exactly the things that marcel mentioned. Blurry photos downscaled are sharper.

But in my case using 20k instead of 40k (point limit custom setting) always gives a better result (with "poor" light/image quality) of camera orientation without downscaling the input data (both are on high).

Cheers

Paul

[Marcel]

Ah, it was in the manual as well, I didn't see that. I just notice the extracted number of points increased quadratically between Low, Medium and High.

The maximum number of points that can be extracted also depends on how many usable points there are in the photo. For photos with a lot of texture we are getting around 200k points on a 21MP image. For 36MP images we've seen values of 350k points or higher. (These are flat subjects with a fine grained surface, so best case scenario).

The cap James noticed at 5000 points at Low setting is probably be the max number of points that can be extracted from the image at that size, and not a limit set by the software.

It's strange that you get better results at 20k points than 40k. I cannot think of any reason why that would happen. From the tests I have did more points always leads to a higher alignment accuracy (but that was with good source photos, I haven't tested with photos in bad lighting)

[ozbigben]

I'd be interested to know the range of Reconstruction uncertainty values that people are getting for 20k and 40k, where the 40k sparse cloud looks worse

[Marcel]

Yes,  and also the actual number of points extracted per photo according to the console. Maybe around 20k is the maximum number of points that can be found, so there is no actual difference between the 20k and 40k setting?
(although the image quality has to be pretty low for that)

[Porly]

Hello,

I had data sets where the alignment with 40k points was not succesful (cameras were not aligned), while with 20k the orientation of cameras was solved. In both situation the maximum set points were found. So with 40k, in each photo 40k points were detected. For 20k, it was 20k as well.

As I already said, the reliability of matches between photos in case of 20k could be higher, so that a threshold value between valid/invalid matches in case of lower amount of points could cause a more robust solution. There are always many missmatches recognized as valid in datasets with poor lightning/quality (noise), so that I can imagine, in case of 20k points per photo, more matches with a better reliability are prefered/used. In case of 40k there could be the same matches, but additionally some of lower quality that could falsify the results.

Thats only my opinion, I found it strange too as I saw better results with lower points. So I thought about it, and thats what makes sense for me.

Unfortunately I am on vacation right now, so that I cant search for a example data set. But when I come back I will try to prepare it for you.