[Note: As with my previous review, I asked 3DFlow for a temporary license to the full version of 3DFZephyr, and they kindly obliged by giving me a 1 month license. This in no way affects the following review.]
I’ve previously reviewed 3DF Zephyr Free (twice!), the free version of this software. The free version has most of the features of the full version, but is limited to 50 photos. Here’s the blurb from their website:
Which limitations does 3DF Zephyr Free have?
3DF Zephyr free can process only up to 50 images, which should be enough for very simple subjects and to play with the first few tutorials. You can also use only one NVIDIA card to speed the processing, and exporting/editing functionalities have been limited (for example, you may only export textures in JPG formats, and so on) and lacks some very important tool (for example, the customized UV mapping tool). However, 3DF Zephyr Free is not time limited, and you may open any .zep file created in any other 3DF Zephyr versions. The full features comparison can be found here.
My Styracosaurus dataset has 53 photos, so testing the free version with that dataset wasn’t really fair. 3DF Zephyr has also seen a lot of development (and as we’ll see later, that’s ongoing), so it seems fair to give it another go, using the full dataset, and try it out on a larger dataset too.
Here’s what you need to know:
Lowest Cost: €149 + Vat (~£158)
(This Lite version has a 500 photo limit – it’s €2400 + vat for unlimited photos, see below)
Time to process Styracosaurus dataset*: 9 minutes
*on default, on my test machine.
A note on that cost: €149 is quite reasonable for a perpetual license for the lite version – in line with Agisoft Metashape, and about the same cost as 4 or 5 months of Reality Capture. But, it does limit you to 500 photos. Not many of my reconstructions use more than 500 photos, and you could conceivably break up large models into chunks and align meshes later, but it does seem a bit arbitrary, especially given that Agisoft Metashape offers unlimited photos in reconstructions for the same price. Still, for people doing ‘normal’ and even ‘large’ reconstructions, €149 isn’t terrible. The license they gave me was for the Pro version; unlimited photos and full editing tools, though I didn’t go over 500 photos in my testing. A full list of variations between versions is available here.
Starting up 3DF Zephyr, the first thing to do is got to Workflow-> new project:
This brings up a window where you can select which parts of the reconstruction you’d like to process. Cameras will be aligned, then you can optionally choose to computer dense point cloud, which if selected will enable you to compute the surface, which enables you to compute the texture. Each subsequent option only becomes available when the previous is selected (for obvious reasons). Each stage can be run manually, rather than through the wizard.
The next screen asks you to select your photos, or – and this is a neat feature – you can select a video and 3DF Zephyr will extract frames for you. Having selected your inputs, you’ll be shown calibration details:
The next few windows allow you to select what style of reconstruction to make – close range, aerial, human body, or urban. I’ve just been using close range because that’s what almost all of my reconstructions are. It’s not clear to me exactly what changes between these settings, though it does recommend Urban even for close-range photogrammetry if photos were taken at differing focal lengths.
You also take a moment to select whether you want to reconstruct on low/medium/high (they refer to high as ‘deep’ for camera reconstructions).
This gets repeated for all the reconstruction options you’ve ticked (cameras, dense cloud, mesh), then you’re brought to a final screen for texture options:
This gives you one final screen summarising your settings, and with a ‘Run’ button at the top.
Hit it, and things start moving:
Here, I’ve enabled the log and system monitor so I (and you) can see resource usage etc.
When the reconstruction is done, you’ll see a summary screen letting your know how many, and which, cameras have been aligned:
And if you dismiss that by hitting ‘finish’, you’ll see your reconstruction:
The reconstruction was really good, as you can see in the close ups below. Very little noise around the horn, complete reconstruction of the base (an issue with some packages), and reasonably good detail with the brass label.
Here’s the sketchfab model:
The texture is not particularly efficiently packed, as can be seen in the image below, but this is something they are actively improving for the next version.
I did also run the reconstruction on deep/high/high, and it took 24 minutes. The resulting mesh had 970,386 faces, vs the 463,292 of default. There was a little more detail, but not a whole lot of difference (high is on the right):
Testing a larger dataset:
Just as with Reality Capture, I ran a 217 image dataset of Neuquensaurus. Unfortunately, with close-range default settings 3DF Zephyr only matched 61 images. That being said, the resultant point cloud was about on par with Reality Capture (which I didn’t make not of how many cameras matched, but it wasn’t all of them):
But, again as with Reality Capture, the mesh was not really usable:
The whole process took about 10 minutes, but that’s not surprising given it only managed to match slightly more cameras than were in the Styracosaurus dataset.
A couple of things that are worth mentioning…
Firstly, 3DF Zephyr Lite and Pro (not free) come with an image masking tool, that’s actually really easy to use. When starting the new project, tick the box ‘mask images’, and then select if you want to use that mask when matching cameras (as opposed to just for the dense reconstruction/mesh):
Then when you click through the project wizard, you’ll reach a page for masking:
Click ‘Launch Masquerade’ and you’ll see a new window for drawing masks:
There’s some really powerful, useful tools in here. In the image above, I’ve made two background strokes (blue) and one foreground (red) and you can see it’s identified the outline of the model pretty accurately.
Here I’ve just quickly used two strokes:
And you can enable a special turntable-based algorithm, and automagically apply the mask to the next image:
This generally works really well when you’re taking photos around an object or using a turntable.
As well as the great masking features, 3DF Zephyr pro also allows scaling via ground control points directly in the software.
And finally, instead of running with the default/low/high presets, you can enable advanced options for each stage:
And mesh reconstruction:
You can probably summize that I’ve been getting on quite well with 3DF Zephyr Pro, but what’s really exciting is what’s coming. 3DFlow make a point of being quite open about the development process, and there have been some interesting blog posts about upcoming features: https://www.3dflow.net/3df-zephyr-development-status-progress-clementine-2/
Very soon, there’ll be an improved UI, generation of normal maps:
And better texture packing:
I really liked using 3DFZephyr – it’s easy to use, the interface is nice, it’s pretty quick, contains a bunch of useful tools, and development is going in a great direction.
The Lite Version at ~£150 is reasonable, but that 500 photos limit nags at me somewhat. It seems quite arbitrary, and isn’t something that affects, say, Photoscan/Metashape. That being said, it’s very rare that I reconstruct a model using over 500 photos, and if I did I could always fall back to Meshroom or COLMAP.
I’m sorely tempted to buy a 3DF Zephyr Lite license for my lab.