Formerly known as ‘Photoscan’, Agisoft’s software was the first of my Photogrammetry Testing series – posted nearly 3 years ago. Here I’m testing Metashape 1.5.3 standard edition. Metashape retains the same competitive pricing of photoscan: $179 for a perpetual license for standard edition. That drops to just $59 with an educational discount, and for that money, there’s no limitations on number of photos or number of GPUs to use, as there is for other commercial packages.
So, what’s changed in that time aside from the name? Well for staters there’s a dark mode to the UI, which is nice, and what I’ll be using for this post.
But more importantly let’s check like-for-like performance, running a batch process.
First thing I do is make sure Metashape is using my GPU(s), by going to Tools->Preferences->GPU:
You can see that I’ve selected my GTX 970 as the GPU to use. The note at the bottom does ask (nicely) that when using a discrete GPU, such as my GTX 970, to turn off integrated GPUS and CPU, so I’ve unticked the intel HD 4600, and the ‘use CPU’ box at the bottom. GPU gets used, as you can see in the image, for most of the stages, so it’s fairly important. Later I’ll have a small comparison between with/without GPUs.
That being set, let’s set-up our batch process, which is much the same as last time
First, I drag my styracosaurus dataset into the empty ‘chunk 1’:
Then I got to Workflow->Batch Process, and in turn I add each stage, applying to all unprocessed chunks:
Then I just hit ‘ok’ and let it do it’s thing:
You can see the GPU getting used for feature matching.
Total time taken: 301.7 seconds
For some reason I cannot fathom, when I did the original review, I gave ‘low’ and ‘high’ settings times, but no medium, so for comparison, I ran the model in Metashape with all settings on ‘Low’ or ‘High’
Low: 95.4 seconds
High: 1226.9 seconds.
For comparison, 3 years ago on the same hardware, the time taken on low settings was 276 seconds, and on high settings was 2706! The speed has more than doubled since then.
The final model on medium was good. It looked very similar to the ‘high’ model I ran 3 years ago, with a bit of noise around the nose:
Metashape actually manages to mesh some of the noise in places, probably due to filtering settings I left as ‘mild’.
However, there are a number of tools present in Metashape that weren’t in the old version of Photoscan I used 3 years ago. Some of the ones that interest me are:
Mesh from depth maps.
The first of these, refine mesh, made the mesh lumpier without really adding much detail:
But the second, generating mesh from depth maps…
…Wow! All the noise is gone, detail is good:
It takes a bit longer to process, but it does use the GPU while doing so, and it doesn’t require that the dense cloud to be calculated at all, so saves time there. But the result is, as you can see above, way better than the old way of doing things by meshing the dense point cloud.
Here’s the final model, on sketchfab:
I also ran the Neuquensaurus through much as I did for the other commercial packages:
3591.92 seconds total, and without a doubt the best reconstruction of this photo set I’ve seen so far.
Other features of note:
A couple of things that I didn’t cover last time but are worth a mention:
You can directly import video now, and Metashape will decompose this to individual frames. Just go to File->Import->Import Video. Useful feature for rapid captures.
Masking tools are pretty good – you can easily mask the background out of images. Most usefully, you can mask a few images, then select points based on those masks, allowing you to clean up a dense point cloud really quickly. Magic wand tools and normal selection tools make things easier, though not quite as quick as 3DF Zephyr with it’s background paint tool.
When I first reviewed Photoscan, it was solid, but some of the free software available really gave it a run for its money. But Agisoft have really made strides – they’ve improved speed, polish, and toolset markedly, and Metashape, like Photoscan before it, remains the most affordable commercial software on the market, particularly with that education discount.
It’s really really hard not to recommend Photoscan to basically anyone. Reality capture might have it beat on speed, but Reality capture costs more per month than Metashape does total (at least with educational discount), and still limits number of photos. 3DF Zephyr is good, but again limits photo number at the same sort of price point we find Metashape standard edition. Meshroom is free, but it’s no where near as quick as Metashape. I’m pretty tight, and I’m fine using Meshroom, but honestly, there’s a clear value proposition in buying Metashape.
thanks for all the great work which saves us other users to save a huge amount of time!
Since I moved to a new M1 Mac there are not so many options for software with good rendering results so I decided to take the commercial way for now.
Your hint to build the mesh from depth maps is worth gold – tests with a photo series which gave my really bad results with “Regard3D” work like a charm with Metashape and your hints.
As a tip to all mac M1 users: Metashape v1.5.5 and v1.6.6 work nice on this little machine but v1.7.3 crashes when building the mesh from depth maps. Building from dense cloud works but the result is not as good.
That’s great to know, thanks – I’m not a mac user myself, so I’d been wondering what works on the new M1 chip. Glad to hear Metashape does, even if the current version is a bit flakey.
Great info, thanks