Setup Instructions

1 Dependencies

  • conda or mamba
  • uv
  • git submodule --init --update to clone the external dependencies (EFM3D is installed from local source)

2 Setup Instructions

2.1 Installation via separate uv envirnoment

cd oracle_rri
# create conda environment for cuda toolkit 12.1
conda create -f environment.yml
conda activate aria-nbv

uv sync --extra dev --extra notebook --extra pytorch3d
# or simply
uv sync --all-extras

This will install all dependencies and create a new python environment in oracle_rri/.venv.

2.2 Optional: xFormers (Linux only)

EFM3D’s DinoV2 attention utilities can optionally use xFormers for memory-efficient attention / nested tensor support. Install it via:

cd oracle_rri
uv sync --extra xformers

2.3 Optional: OpenPoints (Linux only)

# OpenPoints is built via the openpoints-shim dependency during uv sync.
# Ensure the submodule is present (PointNeXt lives in external/PointNeXt).
git submodule update --init --recursive external/PointNeXt

# Default: builds pointnet2_batch + chamfer ops needed by OpenPoints imports.
uv sync --all-extras

# Optional toggles (re-run uv sync to rebuild):
OPENPOINTS_BUILD_POINTOPS=1 uv sync --all-extras
OPENPOINTS_BUILD_SUBSAMPLING=1 uv sync --all-extras
# Chamfer is enabled by default; override only if you want to skip it:
OPENPOINTS_BUILD_CHAMFER_DIST=0 uv sync --all-extras
OPENPOINTS_BUILD_EMD=1 uv sync --all-extras

The build uses the CUDA toolkit found on PATH (aria-nbv provides CUDA 12.1). If nvcc is not found, set CUDA_HOME or OPENPOINTS_TORCH_CUDA_ARCH_LIST explicitly.

3 Data Setup

  • download the Dinov2 ViTB14 checkpoint and place it in .logs/ckpts/
  • PointNeXt-S checkpoint:
    • Google Drive Link - download, and move s3dis-train-pointnext-s-ngpus1-seed1742-20220525-162639-Gz4ViiL95b6KP9MMH2ytG3_ckpt_best.pth to .logs/ckpts/
    • Model Zoo