Auto NuRec — Dockerized Mono-Camera Reconstruction Pipeline

One-shot Docker pipeline for NVIDIA NuRec neural reconstruction from a single (mono) pinhole camera. Runs COLMAP for structure-from-motion and 3DGUT for dense 3D Gaussian reconstruction, then exports USDZ for Isaac Sim.

Based on the official workflow: Reconstruct Scenes from Mono Camera Data.

Requirements

• Host: Linux x86_64 — aarch64 is not supported (e.g. DGX Spark / Grace Hopper). Dependencies including usd-core (USDZ export) and gcc_linux-64 (conda) have no aarch64 builds. Build and run this on an x86_64 machine.
• NVIDIA Container Toolkit (for GPU access in Docker)
• Hardware: NVIDIA GPU (CUDA 11.8 compatible)
• Input: Project folder whose only required content is an images/ subfolder with pinhole-camera photos (e.g. smartphone). You do not need to install or run COLMAP on your host — the image already includes COLMAP and runs it inside the container.

Quick start (images only)

The container has COLMAP and 3DGUT installed. You only need to provide a folder that contains an images subfolder with your photos. The pipeline will run COLMAP (SfM) and then 3DGUT (neural reconstruction) in one go.

# Your folder can be just:  my_project/images/*.jpg
# --shm-size: PyTorch DataLoader needs more than Docker’s default 64MB /dev/shm
docker run --gpus all --shm-size=8g -v /path/to/my_project:/data auto-nurec

All outputs (database.db, sparse/, 3dgrut/ with export_last.usdz) are written into the same project folder.

Project layout

Input: Your project directory must contain an images/ subfolder with pinhole-camera photos.

After the pipeline, the layout matches a typical COLMAP + 3DGUT run (same as running 3DGUT manually with path=<project> and out_dir=<project>/3dgrut):

/path/to/project/
├── images/           # Your photos (JPEG/PNG, pinhole) — required upfront
├── database.db       # COLMAP database (created by pipeline)
├── sparse/           # COLMAP sparse reconstruction (created by pipeline)
│   └── 0/
│       ├── cameras.bin
│       ├── images.bin
│       └── points3D.bin
└── 3dgrut/           # 3DGUT outputs (configurable via OUT_SUBDIR)
    └── <experiment_name>/
        └── <run-timestamp>/
            ├── export_last.usdz   # Import this in Isaac Sim
            ├── export_last.ingp
            ├── ckpt_last.pt
            └── ...

Capture tips (from the NuRec docs):

• ~60% overlap between consecutive shots
• Steady lighting, locked focus/exposure where possible
• Shutter ≥ 1/100 s; avoid motion blur
• Use JPEG or PNG (convert HEIC to JPG if needed)

Build

The image includes COLMAP (apt) and 3DGUT (conda env). Build from this folder:

docker build -t auto-nurec -f Dockerfile .

Run

Mount your project directory at /data (or set PROJECT_DIR to the mount path). The container expects PROJECT_DIR/images to exist. Use --shm-size=8g so PyTorch DataLoader workers have enough shared memory (Docker default 64MB is too small).

docker run --gpus all --shm-size=8g -v /path/to/your/project:/data auto-nurec

Results appear under the project folder:

• database.db and sparse/0/ at project root (COLMAP output; 3DGUT uses project root as path)
• 3dgrut/ (or OUT_SUBDIR) — training outputs; export_last.usdz is under <experiment_name>/<run-timestamp>/ for Isaac Sim

Environment variables

Variable	Default	Description
PROJECT_DIR	/data	Project root (must contain images/)
OUT_SUBDIR	3dgrut	Subfolder name for 3DGUT outputs
EXPERIMENT_NAME	3dgut_mcmc	Experiment name for logs/checkpoints
EXPORT_USDZ	true	Export USDZ for Isaac Sim
COLMAP_MAX_IMAGE_SIZE	2000	Max image dimension for COLMAP feature extraction

Example with custom output dir and experiment name:

docker run --gpus all --shm-size=8g \
  -v /path/to/project:/data \
  -e OUT_SUBDIR=my_run \
  -e EXPERIMENT_NAME=kitchen_01 \
  auto-nurec

Outputs

• export_last.usdz — Under 3dgrut/<experiment_name>/<run-timestamp>/. Load in Isaac Sim: File → Import (or drag into the viewport). Add a ground plane for physics and proxy for shadows; see the NuRec mono workflow.
• Checkpoints (ckpt_last.pt), INPG (export_last.ingp), and iteration outputs in the same run folder.

Troubleshooting

• RuntimeError: unable to allocate shared memory(shm): Add --shm-size=8g (or at least 1g) to docker run. PyTorch DataLoader needs more than Docker’s default 64MB /dev/shm.
• COLMAP “No good initial image pair”: COLMAP produced a weak or empty sparse model. Improve overlap (~60%), lighting, and focus; avoid blur and mixed focal lengths. You can still run 3DGUT if sparse/0 exists, but quality may be poor.
• COLMAP fails / no sparse/0: Check overlap and image quality; ensure enough in-focus, non-blurry images and consistent camera (single pinhole). Try lowering COLMAP_MAX_IMAGE_SIZE if you hit memory limits.
• Out of GPU memory (3DGUT): Use smaller images (e.g. resize before capture or reduce COLMAP_MAX_IMAGE_SIZE) or a GPU with more VRAM.
• CUDA/GPU not seen in container: Use docker run --gpus all and ensure the NVIDIA Container Toolkit is installed.

References

• NuRec — Reconstruct Scenes from Mono Camera Data
• 3DGUT (3D Gaussian Ray Tracing)
• COLMAP