Processing Images

[River123]

Hi,

I am currently using Agisoft Photoscan for my Masters Project. I have a set of 700 images and when trying to process them each stage is taking hours (sometimes days) and then crashing before completing. I tried using low settings and although this worked, the model was very poor quality and defeats the aim of my Masters project of creating a 'high resolution model'. The PC I am using is the best the university has to offer.

Any ideas?

Thanks

[Jeremiah_ROWE]

What sort of imagery is this? UAV?

What is the image resolution? What sort of object/terrain are you trying to model?

[SAV]

Hi River 123,

You should have a look at the system and memory requirements.
1. http://www.agisoft.com/downloads/system-requirements/
2. http://www.agisoft.com/pdf/tips_and_tricks/PhotoScan_Memory_Requirements.pdf

You will have to a) adjust your processing settings to your hardware's capabilities or b) split up the project into smaller chunks, process them individually and then merge them in the end or c) process your imagery on a more powerful machine (or cluster).

Regards,
SAV

[Alexey Pasumansky]

Hello River123,

It would be also helpful if you provide the processing parameters used either from Chunk Info dialog or from the processing report.

[emm]

You will have to a) adjust your processing settings to your hardware's capabilities or b) split up the project into smaller chunks, process them individually and then merge them in the end or c) process your imagery on a more powerful machine (or cluster).

Regards,
SAV

Hi SAV, I like the idea of doing smaller chunks and merging at the end. I am also working on a very large set of imagery.

Which process/program would you recommend for merging at the end?

Thanks,
J

[SAV]

Hi emm,

Merging is done in PhotoScan (Workflow > Merge chunks).
For large projects you would still compute a sparse point cloud (= tie points) based on ALL images, but you would individually run the point cloud densification on smaller chunks.

See this thread for more details of how to do it:
http://www.agisoft.com/forum/index.php?topic=7374.msg35585#msg35585

Regards,
SAV

You will have to a) adjust your processing settings to your hardware's capabilities or b) split up the project into smaller chunks, process them individually and then merge them in the end or c) process your imagery on a more powerful machine (or cluster).

Regards,
SAV

Hi SAV, I like the idea of doing smaller chunks and merging at the end. I am also working on a very large set of imagery.

Which process/program would you recommend for merging at the end?

Thanks,
J

[River123]

What sort of imagery is this? UAV?

What is the image resolution? What sort of object/terrain are you trying to model?

Hi Jeremiah,
Yes I am using UAV imagery which I have obtained using a c-astral Bramor rtk fixed wing. The image resolution is 350dpi

I am modelling a riverscape, so it is a river channel with many paelochannels, and pasture land which is mainly grass. The elevation across the site is fairly consistent also.

[River123]

Hello River123,

It would be also helpful if you provide the processing parameters used either from Chunk Info dialog or from the processing report.

Hi,

It wont let me attach a copy of the report as it is too big but please see below the settings I used.  I used low settings throughout but ideally I need to be using high to get a better resolution model

Processing Parameters
General Cameras 744
Aligned cameras 739
Coordinate system Local Coordinates (m)
Point Cloud
Points 132,742 of 150,256
RMS reprojection error 0.340204 (1.0567 pix)
Max reprojection error 1.02869 (23.5633 pix)
Mean key point size 3.18063 pix
Effective overlap 5.78053
Alignment parameters
Accuracy High
Pair preselection Generic
Key point limit 40,000
Tie point limit 1,000
Constrain features by mask No
Adaptive camera model fitting No
Matching time 3 hours 28 minutes
Alignment time 1 minutes 39 seconds
Depth Maps
Count 739
Reconstruction parameters
Quality Low
Filtering mode Moderate
Processing time 2 hours 42 minutes
Dense Point Cloud
Points 42,155,410
Reconstruction parameters
Quality Low
Depth filtering Moderate
Depth maps generation time 2 hours 42 minutes
Dense cloud generation time 10 minutes 33 seconds
Model
Faces 936,698
Vertices 468,393
Texture 4,096 x 4,096, uint8
Reconstruction parameters
Surface type Height field
Source data Dense
Interpolation Extrapolated
Quality Low
Depth filtering Moderate
Face count 936,786
Processing time 53 seconds
Texturing parameters
Mapping mode Orthophoto
Blending mode Mosaic
Texture size 4,096 x 4,096
Enable color correction No
Enable hole filling Yes
UV mapping time 13 seconds
Blending time 6 minutes 4 seconds
Software
Version 1.2.6 build 2834
Platform Windows 64 bit

[River123]

Hi River 123,

You should have a look at the system and memory requirements.
1. http://www.agisoft.com/downloads/system-requirements/
2. http://www.agisoft.com/pdf/tips_and_tricks/PhotoScan_Memory_Requirements.pdf

You will have to a) adjust your processing settings to your hardware's capabilities or b) split up the project into smaller chunks, process them individually and then merge them in the end or c) process your imagery on a more powerful machine (or cluster).

Regards,
SAV

Hi Sav,

I have considered doing smaller chunks and then merging at the end, but I was worried that when merging the sections together I may lose some quality in my model? It looks like this may be my last resort though. Do you have more advice
on how this can be done accurately?

[Alexey Pasumansky]

Hello River123,

If you are processing aerial survey data and have the coordinate information for the camera positions, then I can suggest to load the coordinate information to the Reference pane and use Reference+Generic preselection option. Also it may be reasonable to increase tie point limit up to 4,000 points.

Also I recommend to update to the version 1.3.3 as it allows GPU utilization for the image matching stage, if there is any graphic card installed on the computer you are using. What is the hardware specification for the system that you are using?

[SAV]

Hi River123,

Here is the workflow:

1. Align all images = generate sparse point cloud. Use REFERENCE in pair preselection (as mentioned by Alexey) to speed up this process. Also add/use your GCPs.

2. When alignment is finished, use Gradual selection (EDIT>GRADUAL SELECTION) to remove tie points which have a large Reprojection error (you want it to be less than 1 pixel, probably somewhere closer to 0.5 pixel. Use the slider to find a 'natural' threshold). Also remove points with a high Reconstruction uncertainty. Again, use the slider to find your 'natural threshold'. Note that Photoscan will only select the points. You'll then have to press the delete selection icon or DELETE on your keyboard.

3. Optimize alignment (magic wand icon in reference pane; use suggested settings)

4. Run the SPLIT IN CHUNKS Python script. Download from here: http://wiki.agisoft.com/wiki/Python

5. Run point densification for each chunk individually. Use batch processing (WORKFLOW > BATCH PROCESS)

6. Merge chunks (WORKFLOW > MERGE CHUNKS)

7. Compute DEM and orthomosaic from dense point cloud of merged chunks. Note that you don't need to compute a mesh if you are only interested in a DEM and/or orthophotomosaic. Saves some time 

8. Be happy about what you achieved 

Maybe consider to upgrade/replace your hardware at some stage (CPU, GPU & RAM). Will make life much easier (no splitting into chunks needed then) 

Regards,
SAV

Hi River 123,

You should have a look at the system and memory requirements.
1. http://www.agisoft.com/downloads/system-requirements/
2. http://www.agisoft.com/pdf/tips_and_tricks/PhotoScan_Memory_Requirements.pdf

You will have to a) adjust your processing settings to your hardware's capabilities or b) split up the project into smaller chunks, process them individually and then merge them in the end or c) process your imagery on a more powerful machine (or cluster).

Regards,
SAV

Hi Sav,

I have considered doing smaller chunks and then merging at the end, but I was worried that when merging the sections together I may lose some quality in my model? It looks like this may be my last resort though. Do you have more advice
on how this can be done accurately?

[River123]

I've attached the computer spec and yes there is a graphics card it is an AMD Radeon R9 270

Hello River123,

If you are processing aerial survey data and have the coordinate information for the camera positions, then I can suggest to load the coordinate information to the Reference pane and use Reference+Generic preselection option. Also it may be reasonable to increase tie point limit up to 4,000 points.

Also I recommend to update to the version 1.3.3 as it allows GPU utilization for the image matching stage, if there is any graphic card installed on the computer you are using. What is the hardware specification for the system that you are using?

[SAV]

Hi River123,

If you want a performance boost, then replace your R9 270 with a GTX 1080.
Here is a comparison: http://gpu.userbenchmark.com/Compare/Nvidia-GTX-1080-vs-AMD-R9-270/3603vs3149

And, as suggested by Alexey, use the latest version of Agisoft PhotoScan which makes even more use of the GPU (i.e., GPU and CPU are now both used in the photo alignment step)

Regards,
SAV