Capturing#
To capture Gaussian Splats, many images from different viewing angles are needed. This can essentially be done with any camera, but some cameras are better suited than others. In my experience, it is easier later on to record a video instead of taking many individual photos.
So here’s a short checklist of what the camera should be able to do:
- a wide-angle lens with as little distortion as possible
- at least 1080p video, ideally 4K
- a high frame rate is not necessary (24fps is plenty)
- ideally optical stabilization (less important in very bright scenes)
For my first tests, I chose a Sony A7 IV full-frame camera with an aspherical Samyang 14mm f/2.8 ED AS IF UMC lens with manual focus.
Training#
A powerful computer is required for training. An Apple M1 or a dedicated GPU with at least 6GB of VRAM (or 12GB and more, depending on scene complexity) speeds up the training process significantly. A strong iGPU can also be sufficient.
Depending on the complexity of the scene, training with this hardware still takes several hours.