Implicit surface renderer for neural view synthesis
Trains a single sample ray -> pixel model by first shifting ray origins along the ray to minimize their distance from the center of the scene
The lastest version will be hosted on Github here.
If you want to modify the code, install the dependencies (npm install
) then npm run dev-server
to start the webpack development server.
- Right click and hold + Mouse / WASD / Space to move camera
- Hold Shift to move faster
- Tab to cycle through view modes (all views in dataset, closest view, predicted view)
- 1 to set your camera to a dataset view
- O to toggle automatic camera orbiting
- R to reset model weights
- T to train for 100 more steps, Shift + T to train for 1000
All views must be loaded to start training. The model will train for 100 steps when all views are loaded.
The lego data is from NeRF, using Heinzelnisse's model