I’ve been trying to embrace the Camera Analysis node in Action but finding it really slow to analyse - recently upgraded my iMac to 2020 version with 128GB RAM and 16GB VRAM and thought I would see some performance improvements over my much older box but . . . a 1200 frame 4K shot has been taking (literally) hours. I’ve set Res to 1 and High Quality for analysis and have ML mattes and Gmasks in the mix which I’m sure are adding to the time but I reckon it will top out at 4 hours by the time interpolation is finished. Admittedly a long shot with all the bells and whistles but I’m curious - is this expected? Is it better to do a roughie at lower quality settings first rather than heading straight for the max?
1 Like
What are the full specs of your iMac?
My first thought was, your numbers don’t surprise me. Machine Learning tools aren’t optimized on the Mac side, and the iMacs aren’t quite the optimized form factor either. For example, similar but different, ML Timewarp on 2019 Mac Pros is cpu only and takes sometimes 30secs/ frame, whilst my Linux box does 2secs/frame.
I’ve been using the heck out of the Camera Analysis on 2019 Mac Pros, both 16 and 28 cores as well as a Linux Threadripper Pro/RTX6000 on a Lenovo and benchmarking on my own. Even my Mac Pros take 2-3 times as long to do some ML functions.
Unfortunately, your machine and its form factor wont shine in these situations. You’ll need to find ways to optimize. And running Full Quality on the first try likely ain’t necessary.
If I were you, I’d find a nice 50 frame 4K clip and spend a day really exploring the new tool, clock some of its operations and renders, and see how to tweak the quality and settings to get what you need.
You can also see if your machine has a comparable entry in the flame benchmark spreadsheet.
Hope this helps.
The way you have set your Camera Analysis will take the most analysis time:
- Res 1 means that every frame in your clip is being used by the algorithm.
- High Quality means that the highest image resolution is being used.
- Using ML Masking adds to the overall analysis time.
For some “easier” shots using a “lower” Res value of 2, 3, 4 or even 5 and a Fast/Intermediate Motion Vectors quality might be good enough. It is a good idea to start with that and see if the results is good enough for what you are trying to achieve.
2 Likes
Thanks guys - that all confirms what I was thinking. The iMac is as tricked out as I could get it but yeah, it’s no Z8 for sure! The solve came out pretty good in the end with just a few tweaks over some dodgy frames, where the old 3D tracker couldn’t find one at all so it’s a win in the end for sure. It’s just so nice to have so many tracking options now as it seems where one won’t get a solve another will.
Thanks to all the devs Fred - it seems like a very powerful new tool. I am very much looking forward to one of these “easier” shots of which you speak 
2 Likes
have you tried down sizing the clip you are tracking? curious to know the speed vs accuracy tradeoff if you were to track the same clip at HD
1 Like
That’s a good idea - i always go for maximum settings as default expecting that that ‘will be better’ I’ll have a look at that when I get some down time.