Crash - how much Ram is recommended? I have 20 gigs

[Brit]

It turned out that the computer wasn't reading 8 gigs of RAM, which explains why it only went up to 24. We pulled them out and put them back in and they worked. Actually the medium setting seemed to work OK. The dense cloud at "high" was 71 million points.

[wojtek]

Another solution is to use 'meshing in blocks' script which allows you to build the geometry in small regions in an automated way. That is only available for the Pro version of Photoscan however. What you can alternatively do is try to do that manually

[Brit]

My new system has defective RAM (and/or a defective Mother Board) that will be replaced tomorrow.

A technician just told me that the main bottle neck now is probably my system hard drive (a 10k rpm Raptor). He said that a solid state drive would make a huge difference –even with 32 gigs ram. Is this true? I just rebuilt my system taking almost 5 days and don’t want to repeat this.

[Brit]

What is (and were do I find) the "meshing in blocks" scripts?

[Alexey Pasumansky]

Hello Brit,

Script samples for the version 1.0.4 and 1.1.0 are published in the following thread:
http://www.agisoft.com/forum/index.php?topic=2734

[StevenF]

Hello Brit,

Script samples for the version 1.0.4 and 1.1.0 are published in the following thread:
http://www.agisoft.com/forum/index.php?topic=2734

I've been using a modified version of this script and it's working great. But the process has taken over a week to run so far.  I made some minor modifications to export the dense cloud of each tile and to save the tile to a new document which can help in saving your progress in case a crash occurs. You can merge the tile documents into a single document using the merge_tiles function listed near the top. Here is the script that works with version 1.1.0 build 1985:

Code: [Select]

## Splits active chunk into user defined grid of sub-chunks (images are kept but bounding box is split)
## after chunk being split, for every smaller chunk mesh is generated based on the dense cloud, then dense cloud is removed from smaller chunks
##additional stripes will be generated and exported

#arguments - X and Y grid sizes

#compatibility PhotoScan Professional 1.1.0
#script v5.3
# Obtained from Alexey at Agisoft on 11/17/14
# Modified by Steven F

# Build dense cloud in main chunk but perform save as after every tile

import PhotoScan
import sys

doc = PhotoScan.app.document
orig_path = doc.path
chunk = doc.chunk


def dist(v0, v1):
"""
Calculate distance between two points defined as vectors
"""
distance = (v1 - v0)
return distance.norm() #**2


# take a folder of photoscan documents and merge them into one document then save
def merge_tiles(doc_tile_folder, merged_file):
import glob
tile_docs = glob.glob(doc_tile_folder + "*.psz")
merged_doc = PhotoScan.Document()
for doc in tile_docs:
tile = PhotoScan.Document()
tile.open(doc)
merged_doc.append(tile)
merged_doc.save(merged_file)

#parts of the model, 3 by default
partsX = 20   
partsY = 20
if len(sys.argv) == 2:   
parts = int(sys.argv[1])
partsX = partsY = parts
elif len(sys.argv) == 3:
partsX = int(sys.argv[1])
partsY = int(sys.argv[2])
print("Script started...")


# dense cloud settings
quality = PhotoScan.UltraQuality
filter = PhotoScan.MildFiltering
reuse_depth = False

# tile doc save and dense cloud export paths
tile_basename = "tileUltra"
las_tile_folder = "/project/ultra_las/"
doc_tile_folder = "/project/ultra_docs/"
las_basepath = las_tile_folder + tile_basename
doc_basepath = doc_tile_folder + tile_basename


# get existing chunk region
region = chunk.region
r_center = region.center
r_rotate = region.rot
r_size = region.size

# part size based on existing chunk region and number of parts
x_scale = r_size.x / partsX   
y_scale = r_size.y / partsY   
z_scale = r_size.z 

# Offset between parts. I don't understand why it wouldn't just be the size,
# but they use center and rotate as well.
offset = r_center - r_rotate * r_size /2.

# Create a dense cloud for each part in a seperate chunk
for j in range(1, partsY + 1): 
for i in range(1, partsX + 1):
print("Starting tile x", i, "y", j)

chunk.label = "Chunk "+ str(i)+ "\\" + str(j)

# set the region for the part
new_region = PhotoScan.Region()
new_rot = r_rotate
new_center = PhotoScan.Vector([(i - 0.5) * x_scale, (j - 0.5) * y_scale, 0.5 * z_scale])
new_center = offset + new_rot * new_center
new_size = PhotoScan.Vector([x_scale, y_scale, z_scale])
new_region.size = new_size
new_region.center = new_center
new_region.rot = new_rot

chunk.region = new_region

# print region center
print("Region center = ", chunk.region.center)

# create a dense cloud for the part
PhotoScan.app.update()
denseStatus = chunk.buildDenseCloud(quality = quality,
filter = filter,
reuse_depth = reuse_depth)

# # export dense points to LAS
if denseStatus == True:
print ("Dense cloud generation successful. Exporting LAS tile")
las_file = las_basepath + "_x" + str(i) + "y" + str(j) + ".las"
exportPoints = chunk.exportPoints(las_file,
  source = PhotoScan.PointsSource.DensePoints,
  format = "las",
  projection = chunk.crs)
print("Point export successful?", str(exportPoints))

else:
print("Dense cloud generation failed. Nothing to export.")

# # save as a new document
new_doc = doc_basepath + "_x" + str(i) + "y" + str(j) + ".psz"
doc.save(new_doc)

# remove depth map, dense cloud, and model to save memory
chunk.depth_maps = None
chunk.dense_cloud = None
chunk.model = None

print("Finished with tile x", i, "y", j)

print("Script finished...")
PhotoScan.app.update()
You'll want to change a number of these settings of course: partsX, partsY, quality, filter, reuse_depth, tile_basename, las_tile_folder, doc_tile_folder

Lately I've been thinking that it would be nice if this process could be sped up by only calculating a depth map once for each image using every image it overlaps. This would probably require a significant change in the way PhotoScan is programmed, but it would remove the need to recalculate depth for an image that overlaps multiple tiles.

[Brit]

I put in 32 GB ddr4 ram, a new processor and a solid state hd and it worked! However I noticed that the scan is skewed so that the rectangular box stele is much higher on one side than the other. Can anything be done to square it? 

[Parker]

Hello,

I bought a fast 64GB data traveller pendrive to use as a readyboost memory. Will agisoft use it as an extra RAM?
 

[pjenness]

HIya

The more ram the better 9obviously) as it avoids any paging etc

My machine at work has 192G and Ive even maxxed that one (but not often)

My home machine has 32G and I can usually keep it running at 24G used then swap if needed)

Photoscan seems to reasonably good at keeping the RAM at a level where you machine is still functional. But at the cost of swapping and slowing down the processing.

Speaking of swap, if you are swapping get a SolidState drive for the scatch disk. It'll help a little at least.

Question to the Agisoft team.
Is there any way on the CPU/RAM intensive processes that you can use GPU ram as first swap? I have 2x 4G 9xx series card, so potentially an extra 8G there not getting used while the CPU processes are running?

Cheers

-P