Forum

Show Posts

This section allows you to view all posts made by this member. Note that you can only see posts made in areas you currently have access to.


Messages - Bzuco

Pages: 1 ... 13 14 [15] 16 17
211
General / Re: Could MS use MORE RAM???
« on: October 15, 2021, 07:54:58 PM »
Alignment phase: 2/3 is done on gpu, so the limit is only VRAM. And overal alignment phase is not so RAM/VRAM hungry, so I guess no reductions in memory consumption were applied here at the expense of performance.

The depth maps generation phase is done on gpu and you need to feed GPU through PCIE slot(which is fast enough, so no bottleneck).
You can check during this phase if there are some reads from disk and if, what are the values in spikes, if just few MB/s or tens or hundreds of MB/s. If no significant reads, everything was in RAM  :P .

Next phase is filtering depth maps. In this phase RAM usage depends on few factors. How many depth maps you will filter together(BuildDenseCloud/max_neighbors parameter in tweaks) and of course what is the resolution quality.
Next more important factor for RAM requirements is what is the shape of the object you were photographing and where you were taking photos.
In scenario with one long wall and camera positions one next to each other pointing in the same direction...this is not RAM hungry, because there will be just few(3-4?) depth maps needed together in filtering process at the same time.
In scenario when you were orbiting around object and you take several hundreds of photos pointing in the ~same direction...this will be very hungry, because lot of depthmaps need to be filtered together at the same time.
It would be difficult to predict what scenario might occur and make MS smart enough in terms of RAM decision...if keep more data in RAM or temporarily put data on disk.

This is how I see RAM requirements...more in depth analysis only with MS programmers :D

The last phase of dense cloud generation is joining and storing final data on disk. Here you can speed up things if your project is on SSD instead of classic HDD. I am using HDD and often see, that MS is waiting on disk to write all dense data, but in the size of my projects it still doesn not matter. Your projects are different story :D.

212
General / Re: Could MS use MORE RAM???
« on: October 15, 2021, 04:58:00 PM »
Windows is automatically keeping last used files in RAM memory until memory is not needed for other heavy tasks. So in scenario you are copying photos from GoPro somewhere on disk, these files are automatically in RAM memory untill restart computer. And if you right after copying run MS and start processing, those files are reading from RAM. That is how windows automatic caching works and programming languages when are asking for some files from disk(first there is still lookup in RAM, and if not there, then read from disk).
If you come to processing in MS after restart PC, you still need to read photo files from disk first time, but any other following work will be using copies of those files in RAM.
Even when MS is creating some files during process and storing them in zip files on disk, those files are immediately in RAM.

You can check in task manager different memory values, total, available, cached, free. RAM memory is used more than just actual usage value.

213
General / Re: 100% CPU for minutes after loading heavy project
« on: October 15, 2021, 12:08:52 PM »
The latest 1.7.5 13229. In atachment how it looks after opening. Second chunk is active in tie points view mode.
The same behavior in previous 1.7.4 13028 and older.

214
General / Re: 100% CPU for minutes after loading heavy project
« on: October 15, 2021, 10:15:38 AM »
+1

I also noticed this behavior. When I checked resource monitor, MS just after opening project is reading all depth maps zip files from project sub folders. In my case <30s, 200 cameras and 8 core cpu.

215
General / Re: Project file size - Orthophoto
« on: September 29, 2021, 01:31:02 PM »
I am not sure, but I think, that orthomosaic file is internaly in uncompressed format, like simple bitmap 8bits/channel.
For example 18Mpix jpeg from camera has around 10MB in highest JPEG quality, it's bitmap equivalent is 51MB... 5 times more. If you make some overlap when taking photos, then 3.2 times bigger orthomosaic seems to be correct and minimal size.

216
General / Re: Multiple GPU Priority & and VBO settings
« on: September 27, 2021, 04:48:56 AM »
What you probably see in monitoring is just simply huge difference in computational performance between those cards (3090 is ~13x more powerful than M4000).
Assigning task and feeding GPU with data takes some time and if metashape is assigning regularly same amount of data to each card then 3090 computes it much faster and is more in idle time(or idling during transfering next data) than M4000.
You can run gpuz utility twice and monitor both cards more detailed and check what was gpu utilization %, memory allocation, memory controller %,  ...at the same time.
If your motherboard is able to keep speeds of both PCIE slots the same when utilizing two cards(description in motherboard manual), then it should not matter the order of cards. But I would still swap positions, because you will not have M4000 forever.

217
General / Re: factors that affect speed of progress
« on: September 26, 2021, 02:05:53 PM »
If you want to make some test run before final one and do not want to wait hours for the result I would advise:

In align photos phase:
-keep accuracy on high because we still want as much precise pointcloud as possible. Medium acc. make sense only if you are sure, that your photos are very good quality and feature rich... otherwise med. acc. could cause errors in camera positions or completelly bad camera position especialy if photos are not perfect.
I would set keypoint limit to 5000 for interior projects(or where all objects on photos are near you) and higher values 10k-20k for exterior projects(where objects are few meters near and also houndreds of meters far from you, e.g. mountains)
With generic preselection enabled, whole alignment phase takes only fraction of total project computation time, so no need to experiment with parameters to save few seconds/minutes...

In build dense cloud phase:
-also here you can keep the quality to your final one value(depth maps will be computed only ones and not twice for test run and for final run).
To speed up test run and see if the result is good and expected it is useful set tweak parameter in preferences dialog: parameter "BuildDenseCloud/max_neighbors" with value 10 or 20. The result will be more noisy or better said with less prefiltered points but computation time will be 2 - 3x quicker and you will still see if the pointlcoud is as expected or with some completelly misaligned parts or doubled parts.
For final run depth maps will be reused from test run and it is enough set the "BuildDenseCloud/max_neighbors" to its default (if I remember 50) or set it even higher...it depends of your projects, how many photos are overlaped and what was your object of interest.

I am only using photogrammetry in garden projects(house, backyard, garden, road) but this debate should be splitted in more parts, because making photogrammetry is slightly different for interiors, exteriors, top down aerial photos with drone, objects on turn table, human body/face scanning, etc.

In every case it is good advise to have enough photos(no more or no less than it is necessary) taken from good positions/angles and with best quality as possible.
I hope you will find something useful in my post.

218
General / Re: Tweak to limit GPU load during dense cloud generation
« on: September 23, 2021, 11:08:07 AM »
You can download MSI afterburner util and set power limit to what value you need, eg. 70%. The power limit range depends on GPU and how vendor set that in bios.
Lowering power limit will decrease automaticaly core clock, so you can increase it back using GPU core clock slider...in my case I am able with 70% power limit increase core clock max. +120MHz and it is stabel. Find your best value from +0MHz and increasing it with +20MHz steps. ...This is how you can run your GPU with lowest voltage at some frequency and also with much lower temperature.

If you have RTX card, then you can additionaly use nvidia driver utility with this command:  "C:\Program Files\NVIDIA Corporation\NVSMI>nvidia-smi.exe -i 0 -lgc 300,1000"
-i 0   is index of your GPU, first one.
-lgc  will set minimum and maximum core clock frequency, in my example min 300MHz and max 1000MHz
- with this util you have more control on your GPU

219
General / Re: Do i need a "sli" motherboard to use 2 gpu ?
« on: September 20, 2021, 04:59:54 PM »
RTX cards does not use NVLink SLI Bridge, except 3090. You can use any motherboard with 2 or 3 PCIE slots.
Metashape is transfering data to and from GPUs through PCIE slots, bridges are not needed.
PCIE slot speeds are also not important, because even PCIE 3.0 1x can transfer ~1GB/s which is enough.

220
General / Re: double geometry
« on: September 16, 2021, 08:50:32 PM »
Show cameras icon from toolbar. I always check camera positions, if they are ~where I was taking photos...sometimes it happens to me, that one/two cameras are aligned, but are underground, or outside of area etc.

221
General / Re: double geometry
« on: September 16, 2021, 08:31:57 PM »
Maybe a few cameras are a little misaligned?
From my experience, metashape can handle even shadow/sun areas pretty well, until shadows are not too dark.
Can you check visually if all cameras positions are on drone flight trajectory, or some of them are slightly off the trajectory?

222
General / Re: sparse cloud clean, dense cloud mess
« on: September 13, 2021, 07:42:48 AM »
Yes, sparse cloud consists from points with different values of errors. You can see it if you start select points using gradual selection sliders. Points with higher error values can potentially cause deviations in camera positions/rotations, which could later results in deformations in shape of dense cloud/mesh.

223
General / Re: Nvidia Tesla K80. A good, cheap option?
« on: September 12, 2021, 08:16:41 PM »
According number of cuda cores, core frequency, boost frequency and total GFlops power you can expect performance ~exactly to your 2070super.
It is questionable, if CUDA 3.7 compute ability is enough for metashape.
With OpenCL 3.0 should be no problem.
It is a little bit power hungry...300W

224
General / Re: sparse cloud clean, dense cloud mess
« on: September 12, 2021, 11:01:34 AM »
From the PDF manual: The sparse point cloud represents the results of image alignment and will not be directly used in  further  processing...
Delete sparse points make sense only using gradual selection function(from Model menu), which will results in better(more precise) camera positions. After selecting and deleting points using gradual selection tool, you need to use optimize cameras function from Tools menu to update camera positions.

If you want to clean dense cloud points, you can select them by color, by Mask, or filter them by confidence.(Tools-Dense cloud menu). Or you can export dense cloud to Cloud Compare program and apply SOR function which will remove points representing noise in cloud.

225
General / Re: New Low Hash Rate (LHR) GeForce Cards
« on: September 06, 2021, 10:07:06 AM »
Despite my concerns about dispatcher in GPU, 3060ti is doing great job after see your more detailed results.
I am INTEL user, but from some posts on internet 5600x users without PBO enabled have low cinebench multicore scores, just ~4.07GHz instead ~4.6GHz. https://linustechtips.com/topic/1269138-5600x-low-cinebench-r20-scores-multi-stock/   Can you check your frequencies during dense cloud phase?
I would not expect suc big difference in dense cloud times, it is more than twice. From my perspective the only reason would be that 9900K was pretty high clocked, and 5600x is keeping lower multicore freqv. for some reason?...PBO, power limit, temp limit?

According review from techpowerup, 5600x is handling max freqv. great in terms of different instructions utilization https://tpucdn.com/review/amd-ryzen-5-5600x/images/boost-clock-analysis.png
9900K https://tpucdn.com/review/intel-core-i9-9900k/images/clock-analysis.jpg

Pages: 1 ... 13 14 [15] 16 17