A month ago I finished my master’s thesis about visualizing large scale LiDAR datasets in Unity and the workflow to get there (“3D VISUALIZATION OF GEOLOGICAL STUDY AREAS FROM HIGH‐DETAIL GEOSPATIAL DATASETS USING UNITY”).
I’d like to share some of my experiences here as I used RealityCapture to mesh and texture a dataset with over 1.9 billion points plus several hundred additional images and export it to Unity.
Altogether, that worked out nicely. The final (walkable/flyable) model spans over 300.000 m² and the final executable Build takes up 13 GB of space, with an initial dataset of over 270 GB. Some selected points of interested in the model show textural details smaller than 1 mm and geometrical details of around 1 cm.
I had the laser scans registered beforehand, so RC was only used to reference the additional pictures and create the models.
Automatic image referencing proofed very fast and reliable. However, the areas were I had to meticulously assist the referencing often showed texture-artifacts in [*]the final version.
For several reasons I couldn’t just let RC mesh the entire dataset at once and had to split the model into parts using the reconstruction regions. This was quite laborious, but delivered a grid of differently detailed model with up to three Levels of Detail.
Export into Unity:
The engine accepted the fbx output without complaining and allowed for a performant visualization of the dataset after a few tweaks.
The final version, after some heavy simplifications and some optimization, was usable on a medium desktop pc with over 30 fps and made use of several interactive features.
Encountered issues:
- Model borders had to be cut off to get a seamless transition
- Modelling/texturing often failed at high detail due to the limited RAM of 32 GB.
- Exported models often showed stray polygons at the edge of reconstruction regions.
Best feature: Simplify
The process was incredibly quick and allowed me to create several LOD of every tile in the grid for Unity.
Wished for UI features:
- automated gridding instead of manual setup of reconstruction regions
- simple reconstruction presets for these grid-tiles to create several LOD
- Batched processing of such a grid
Criticism: Documentation
I never really got the hang on the documentation. I found it quite awkward to use and difficult to remember where to find what information. In many cases I had to rely on trial and error as I couldn’t find the information on what a setting was doing. Additionally, it being part of the interface often made it difficult to read while using the software. A PDF or website would have been more useful.
Conclusion
RealityCapture proofed very useful to me and didn’t have many of the limitations of the other solutions I’ve tried when dealing with such a dataset. The documentation made my life a little harder than necessary as the software could be quite easy to learn and efficient to use. I’d like to revisit my master project in the near future and again use RC (with my additional experience) to get the LiDAR set modelled and textured. I’m looking forward to future developments
KleinC3.png
KleineKarte.png
ResultMap.png