[Agisoft Photoscan Pro version 1.2.6] Why is my point cloud render taking so long

daniellooartist

Issue: Taking over a day to create a simple model. Every tutorial I have seen has taken 40 minutes to an hour. Either their computers are from the year 4825 or I'm doing something wrong. It's most likely he latter. Granted my resolutions higher but my bake is taking over 24 hours. That's beyond disproportional. I admit I was playing the Witcher 3 while it was rendering, but that's because I waited 5 hours prior and the bar wasn't even halfway. Sorry but I'm not going to let my CPU get locked down for an entire day.

Steps Taken To Resolve
>Using 80 photos
>Resolution 7360x4912 (1.77MB per photo)
>Optimized scene.
>Re-exported my high, low, and cage.
>Baked other maps. Nothing wrong with my other map bakes. JUST my color.
>Align Photos: Accuracy: Highest
>Align Photos: Pair Preselection: Disabled
>Align Photos: Pair Key point limit: Points: 40,000
>Align Photos: Pair Tie point limit: 4,000
>Align Photos: Constrain Feature By Mask: Off
>Align Photos: Adaptive camera a model fitting: On

>Dense point cloud: Accuracy: Medium
>Dense point cloud: Depth Filtering: Agressive
>Dense point cloud: Reuse depth maps: off

Results: Pathetic! This is my DENSE point cloud which took over 20 hours.



Specs
Maya Version: Agisoft Photoscan Pro version 1.2.6 build 2834 (64bit)
OS: Windows 10 Pro (Build: 14393.447)
Video Card: Nvidia GeForce 980 GTX (Driver: 373.06)
CPU: Intel Core i7 3930K

AlecMoody

That dense cloud is waay too sparse. What does your result after align photos look like? My guess is your photos are not good enough and photoscan is having a hard time matching points. Also, make sure you are deleting all inaccurate or unneccesary points at each step in the process.

I would suggest you bake your final game models in something other than photoscan.
Handplane baker doesn't currently support baking highpoly textures but it is on the roadmap to be added. Xnormal might be your best bet, but is slow at handling really high poly models.  If your mesh is dense enough, you could convert your textures to vertex colors in something like zbrush, then bake those.

daniellooartist

The photos were fine. I know that wasn't the problem. The distribution was even. I've seen people take far fewer photos at significantly lower resolutions and bake a much better model with far less time. Here are 3/80 random ones to make my point.

Scruples

80 photos is too much for this model imo, I would take a photo every 15° and then another batch higher up every 30° which should net you around 36 photos. I'm guessing what happened was your computer ran out of memory and started using the pagefile which is going to really slow things down.

You point cloud should be more dense, even at medium, because it's fairly sparse it shows that it's having trouble picking accurate points, I believe this is caused from using a flash as your primary light, it's ok to use it as a fill but not as your primary source. I find it the easiest to go outside on overcast days if I'm going to photoscan anything. Also your aperture is still a bit wide and you're losing a lot of detail from dof.

You're shooting using an a7r? 36megapixels x 80 puts you in the 240-photos range.
(from their memory requirements pdf http://www.agisoft.com/pdf/tips_and_tricks/PhotoScan_Memory_Requirements.pdf)
•Building Model (Arbitrary mode)
200x12mp photos - Medium Quality 6-18gb memory required.

You can change the size your jpegs in-camera, try medium or even small.
http://www.peachpit.com/articles/article.aspx?p=2201797&seqNum=2

daniellooartist

Now that I think of it, Moody may have been right. Is this a good coverage or does there need to be more overlap? Also, can I just take twice as many photos with my iPhone camera and get the same resolution? The reason I ask is because the 36MP camera burns out easily.

AlecMoody

The flash is your biggest problem. More overlap and a more organized approach to how you circle the model would be better. If your shooting hand held and don't have access to lighting, shoot outdoors in open shade on the brightest day you can. You might find that once you stop down enough to keep everything in focus you shutter speed is too long to hand hold. Hopefully your camera has good high iso performance or you can get a tripod.

daniellooartist

Does this look ok for a base point cloud? I did this one with my iPhone camera. 96 pictures.  I don't have any area in my house were lighting was even so I left flash on. But the iPhone6's flash isn't too brutal. I THINK I have good overlap. The reason I'm asking before I finish the dense point cloud is because I don't think it's going to be finishing any time soon.

daniellooartist

Are you freaking kidding me!? This is my dense cloud render. It looks almost no different than the basic cloud render. I had it set to medium accuracy and miled but how could it get this bad!! Does this even make sense or is it just me?

daniellooartist

wait, never mind. I didn't know I had to switch veiws.

AlecMoody

So what does your dense cloud look like?

You should really get rid of the on camera flash. The software has to find matching features in order to match points. When you use a flash you are moving the lighting in each image causing the colors in the images to change. Basically, in the center of the object you have a highlight that moves, that part is difficult for photoscan to match. On the edge of the model you have heavy shadows with no detail, again difficult for photoscan to match.

oglu

you could use a pole filter to get less reflection...
https://docs.unrealengine.com/udk/Three/TakingBetterPhotosForTextures.html

daniellooartist

 I think you're right. I could get rid of the flash. It gets darker earlier this time of year so I guess I'll have to be quick.

ojkoorde

Old post but anyway ... someone may find it helpful. You must keep your light consistent among all photos. Every pic you take has different light conditions due to flash. How is Photoscan supposed to figure out common points between photos if they have different color intensity so to speak. 2nd: jpg and a cellphones are not what you use for photogrametry. You should use image format which keeps all the original info from the image sensor. Jpg uses lossy complession making things even harder for Photoscan. Moreover your pics should be lens distorted, have removed vinegetting and chroma aberration to be most useful. All this can be achieved using some decent camera and photoshop for lens distortion correction from a raw format provided (not jpg). Also to help Photoscan your pics should have lens focal lengh info embedded which jpg does not provide. You must also use small aperture to keep as much as possible in focus. Last - use small ISO to keep noise off.

TTools

I'll add one more note to this, though I imagine it may be common knowledge. Just covering bases for anyone else who comes across this post. Regarding extremely long processing times, make sure you enable OpenCL in the preferences tab.  You'll need to set cores used to one less than total so that your GPU can be used.  This will exponentially reduce calculation times.  More details can be found here:
http://www.agisoft.com/forum/index.php?topic=6193.0