Gamegoof, I feel your pain with how your (previous) company approached photogrammetry
Having said that, I honestly believe Photoscan is not to blame here, what it does, it does wonderfully and last I checked (granted it was a while ago) it allowed A LOT more flexibility and a lot more tools for real artists than Autodesk solution. Surely Autodesk will not sit idle and will continue improving, but it won't spit out game-ready assets anytime soon, if ever.
You mentioned artists with no previous experience with photogrammetry cleanup, and with no previous experience with retopology/resculpting/baking, failed on their very first attempt. That is not surprising at all, and is hardly Photoscan's fault. I don't mean to be harsh, just trying to be fair with respect to Agisoft.
Let me break down main issues with photogrammetry-based pipeline, assuming person taking the photos did fantastic job and you have great highres source mesh:
- noise in mesh: unavoidable in real world, if you are going for very detailed mesh, you will get noise. It's not that hard to smooth out whatever needs smoothing, but surely requires some practice and, well, time. Alternatively, you can process scans at lower quality which produces much smoother results but retains a lot less detail.
- very high polycount and 'messy' triangulated mesh - I don't know of any software that automatically creates good topology and doesn't waste a single polygon. Some auto-retopology tools appear lately but all they do is generate 'pretty' regular topology, which may be of some use to animators, but for games you need adaptive decimation of your geometry (keep geometry only where you really need it, reduce polycount everywhere else), so that's really no good. I just don't see any alternative to manual optimization/retopology. No automatic solution will know where you need detail and where you don't.
- efficient UVs, this is something that Photoscan could improve, but there are so many UV solutions on the market that it's often more convenient to stick to one that you use on a daily basis.
- texture contains 'shading', by far the most problematic issue to overcome, unless you just want to keep geometry and create texture from scratch, or just don't care about in-game lighting/shaders. Few examples of this: any crack in the scanned rock contains 'shadow' in texture, and then lighting in engine puts shadow on top of that shadow, which results in unrealistic ubershadow
If you get specularity or reflection, or Fresnel, in your scan-generated texture, it will appear static, while in reality those things change relative to viewing angle/light vector. All these things are just there in photos - this has very little to do with Photoscan. PS actually does try to reduce some of these issues (with enough photos from different angles, it will significantly reduce specular reflections on human face, etc.).
- geometry and texture is unique (non-tiling), which is great, but these assets eat up a lot of resources. I wouldn't recommend large-scale use of photogrammetry for games that need to run well/stream smoothly on previous gen consoles etc. Again, not something that Photoscan can or even should attempt to fix.
Games were always about a lot of work to create something that looks great and yet renders in fraction of a second. This doesn't change with photogrammetry - you need lots of experience and smarts in order to make use of scans, it's just that we collectively lack experience. But we'll get there! The really sad part is that big companies place so little value in research and learning of new tools and workflows...
-Andrew
Show Posts
