Skip to content

Latest commit

 

History

History
56 lines (41 loc) · 2.63 KB

README.md

File metadata and controls

56 lines (41 loc) · 2.63 KB
Raw

Nerfstudio fork for Stanford NAV Lab

This is a fork of the Nerfstudio repository. It is used in the construction of Neural City Maps and Neural Elevation Models, projects by the Stanford NAV Lab.

Installation

Follow the instructions for Nerfstudio installation here up to "Dependencies" (create conda environment and install PyTorch and dependencies). Afterwards, clone the repo, switch to this branch, and install the nerfstudio package from source:

git clone https://github.com/Stanford-NavLab/nerfstudio.git
cd nerfstudio
git checkout adam/terrain
pip install --upgrade pip setuptools
pip install -e .

Then, install the terrain_nerf package:

cd terrain_nerf
pip install -e .
ns-install-cli

Data

We use simulated imagery from Google Earth Studio as well as real-world aerial drone imagery. Drone imagery is available at https://dronemapper.com/sample_data/ (we use the Red Rocks, Oblique dataset).

Google Earth Studio (GES)

GES allows for rendering imagery of any location on Earth (and the Moon and Mars) using Google Earth. An account is needed.

  1. Generate GES .esp project file from lat/lon/alt using scripts/generate_ges_traj.py
    • e.g., python scripts/generate_ges_traj.py ges_traj.esp 37.333976 -121.8875317 200 --template Templates/Transamerica.json
  2. Load the .esp file into GES (create blank project, then import)
    • Set total time to 10 seconds (with 30 FPS for a total of 300 frames), check "Scale existing keyframes"
  3. After setting settings, then render.
    • Drag Google Earth logo to bottom right corner
  4. Use scripts/ges2transforms.py to generate transforms.json.
    • e.g., python scripts/ges2transforms.py ../nerfstudio_ws/GESSanJose/ san_jose.json 37.333976 -121.8875317

Preparation

Data preparation is identical to that of Nerfstudio.

  1. Create a /data folder within the repo.
  2. For each scene, create a folder within /data (e.g., /Scene01).
  3. Inside the scene folder, place imagery and a transforms.json file containing camera poses and parameters. If needed, use COLMAP or Nerfstudio's ns-process-data to estimate camera poses.

Training

As per Nerfstudio training procedure, run the following command:

ns-train terrain-nerfacto --data data/Scene01

and monitor training through Viser and/or Weights and Biases.

To save height field weights, use scripts/save_nemo_weights.py.