VRoom

VRoom is a comprehensive 3D reconstruction and rendering pipeline that generates virtual rooms from scenes. The pipeline processes raw scene images through Structure-from-Motion (SfM), semantic segmentation and tracking, Gaussian Splatting training, and final mesh generation. The repository also includes a custom-made differential 2DGS rasterizer and a companion mobile application.

The full pipeline is coordinated via full_pipeline_runner.py, which executes the following stages:

1. SfM (Structure-from-Motion): Camera pose estimation and sparse point cloud generation via COLMAP.
2. Masks & Tracking: 2D semantic segmentation (using SAM3), object tracking across frames, and 3D voting for label consensus.
3. Gaussian Splatting Training (gstrain): 2D Gaussian Splatting training on the reconstructed scene using a custom-made CUDA differential rasterizer.
4. Object Refiner: Optional Per-object quality enhancement by selecting the best real observation, generating a novel-view orbit with SV3D, and training a dedicated per-object Gaussian model from the combined real and synthetic views.
5. Mesh Generation: Extraction of RGB, depth, and semantics to generate final 3D meshes for objects.

• full_pipeline_runner.py: The main script to run the entire end-to-end pipeline.
• sfm_label_runner.py: A combined runner for only the SfM and Semantic Labeling stages.
• sfm/: Structure-from-Motion utilities using COLMAP.
• masks_and_tracking/: Segmentation and object tracking scripts.
• gstrain/: Gaussian Splatting training module.
• mesh_generation/: Utilities for extracting meshes from the trained model.
• object_refiner/: Per-object Gaussian refinement using SV3D novel-view synthesis.
• diff-surfel-rasterization/: CUDA rasterizer backend.
• Mobile-APP/: React Native (Expo) companion mobile application.

git clone --recursive https://github.com/EzzCode/VRoom.git

If you cloned the repository without --recursive, pull the necessary submodules first:

git submodule update --init --recursive

The required submodules are defined in .gitmodules

Before building and running the pipeline, ensure you have the following installed and configured on your system:

• CUDA Toolkit: Required for the CUDA rasterizer and GPU acceleration. Download and install the appropriate CUDA Toolkit version for your system. Ensure that the CUDA binary directory (e.g., C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\bin on Windows or /usr/local/cuda/bin on Linux) is added to your system's PATH environment variable.
• COLMAP: Required for the Structure-from-Motion (SfM) stage. Ensure the directory containing the colmap executable is added to your system's PATH.
  • Windows: Download the COLMAP binaries.
  • Linux / WSL: We provide a helper script to build COLMAP 3.10 from source with CUDA enabled. Run ./environments/install_colmap_linux.sh from the repository root.
• SAM3 / Ultralytics: Required for the Masks & Tracking stage. The ./sam3.pt segmentation model weights are downloaded automatically by Ultralytics on first use (it will be saved to your current working directory). Set the --ultralytics_home argument (or the ULTRALYTICS_HOME environment variable) to a directory with sufficient storage space for the checkpoint cache. A CUDA-capable GPU is strongly recommended.
• SV3D (sv3d-diffusers): Required for the Object Refiner stage. Clone the chenguolin/sv3d-diffusers repository into external_deps/sv3d-diffusers/ inside the workspace root:

  git clone https://github.com/chenguolin/sv3d-diffusers external_deps/sv3d-diffusers

  Model weights are downloaded from Hugging Face on first use. Set the HF_HOME environment variable (or update HF_CACHE_DIR in object_refiner/constants.py) to point to a directory with sufficient storage space.

To build the custom CUDA rasterizer required for Gaussian Splatting, the method depends on your operating system.

For Linux:

make clean build

For Windows:

cd diff-surfel-rasterization
python -m pip install -e . --no-build-isolation

The pipeline requires specific Conda environments to run different stages:

• masks: Used for the Masks & Tracking pipeline.
• pipeline: Used for SfM, Gaussian Splatting training, and Mesh Generation.

You can create these environments using the provided configuration files in the repository:

For Windows:

# 1. Create the masks environment
conda env create -f environments/environment_masks.yml

# 2. Create the pipeline environment
conda env create -f environments/environment_pipeline.yml

For Linux / WSL:

# 1. Create the masks environment
conda env create -f environments/linux_environment_masks.yml

# 2. Create the pipeline environment
conda env create -f environments/linux_environment_pipeline.yml

To allow the runner scripts to properly import the core modules (gstrain, masks_and_tracking, etc.) across the pipeline, install the repository as an editable package in your Conda environments.

Activate your environment and run from the repository root:

pip install -e .

(This uses the pyproject.toml file to manage package imports and dependencies).

You can run the entire pipeline end-to-end using the full_pipeline_runner.py script:

python full_pipeline_runner.py --data_path /path/to/scene_folder

Key Arguments:

• --data_path: Path to the scene folder containing an images/ subdirectory.
• --out_base_dir: (Optional) Specify a unified directory to save all pipeline outputs.
• --skip_colmap, --skip_masks, --skip_training, --skip_mesh_gen: Flags to skip specific stages of the pipeline.
• --dry_run: Print the commands that would be executed without actually running them.
• --small_run: Limit training iterations and frames for a quick test run.

For a full list of configuration options, run:

python full_pipeline_runner.py --help

If you only want to run the COLMAP reconstruction and semantic labeling, use:

python sfm_label_runner.py --data_path /path/to/scene_folder

After full-scene Gaussian training, run the object refiner to produce per-object high-quality Gaussian models. The refiner requires the trained scene model path and the scene output directory:

python refine_all_objects.py \
  --model_path /path/to/training/gs_model/<run_timestamp> \
  --scene_dir /path/to/scene_output \
  --iterations 20000

Key Arguments:

• --model_path: Path to the trained scene Gaussian model directory.
• --scene_dir: Root output directory of the pipeline run (contains labeled_output/, tracking_output/, etc.).
• --object_ids: (Optional) Space-separated list of object IDs to refine. Refines all objects if omitted.
• --iterations: Number of training iterations per object.
• --reuse_sv3d: Skip SV3D generation if outputs already exist on disk.
• --debug: Enable visual debugging artifacts.

For a full list of options, run:

python refine_all_objects.py --help

If there is a specific make target for training, you can use:

make run_train DATASET=3dovs

(Note: Ensure you have the pipeline environment activated when running standalone training commands).

The VRoom 2DGS pipeline can be deployed as a serverless backend using Modal. This allows you to offload the heavy CUDA-accelerated processing (COLMAP, SAM3, and Gaussian Splatting) to an A10G cloud GPU while exposing a clean REST API.

Detailed instructions on how to configure, deploy, and interact with the backend API are not included in this file. Please refer to the dedicated backend guide:

👉 Vroom_BE.md

The repository includes a mobile application located in the Mobile-APP/ directory. It is built using React Native and Expo. To run it, navigate to the directory and install dependencies:

cd Mobile-APP
npm install
npx expo run:android