BayesFusionSDF

BayesFusionSDF provides research code, configuration, and reproducibility utilities for our paper.
A lightweight project page is hosted at: https://bayesfusionsdf.soumyamazumdar.com
Main website: https://soumyamazumdar.com

Repository policy (important): This repository does not redistribute third-party datasets (e.g., CO3D or other restricted/terms-governed datasets). Instead, we provide scripts and clear instructions to obtain datasets from their official sources and to generate the required preprocessing outputs.

Overview
Project Page (GitHub Pages)
Installation
Data
Training
Evaluation
Reproducibility
Directory Structure
Citation
License
Contact

Overview

This repository contains:

core implementation (models / pipeline / utilities),
configuration files for experiments,
scripts for dataset preparation, training, and evaluation,
documentation and a GitHub Pages project page.

Replace this section with a short formal abstract of your method and the primary contributions.

Project Page (GitHub Pages)

The project page is served from the /docs directory via GitHub Pages and is configured for the custom domain:

bayesfusionsdf.soumyamazumdar.com

Installation

Option A — Conda (recommended)

conda create -n bayesfusionsdf python=3.10 -y
conda activate bayesfusionsdf
pip install -r requirements.txt

Option B — pip + venv

python -m venv .venv
# Linux/macOS
source .venv/bin/activate
# Windows PowerShell
# .venv\Scripts\Activate.ps1

pip install -r requirements.txt

If your code requires CUDA / PyTorch versions, specify them explicitly in requirements.txt or provide a separate requirements-cuda.txt.

Data

Policy

This repository must not include:

full copies of third-party datasets (e.g., CO3D),
any dataset files that violate licensing / terms,
any personal or sensitive data.

Instead, include:

a download guide (links to official dataset instructions),
a preprocessing script that produces the exact directory layout expected by the training code,
a small toy/example sample that you created or that is permissively licensed (optional).

What you SHOULD provide in GitHub

Provide instructions + scripts, not the dataset.

Recommended:

scripts/download_<dataset>.sh or scripts/download_<dataset>.py
- downloads from the official source (or prints the official steps if manual acceptance is required)
scripts/preprocess_<dataset>.py
- converts raw dataset to your required format
configs/data/<dataset>.yaml
- describes expected paths, splits, and preprocessing options
DATASETS.md (recommended)
- a single page listing supported datasets, licenses/terms pointers, and step-by-step setup

Example dataset layout (you define yours)

data/
  <dataset_name>/
    raw/
    processed/
      train/
      val/
      test/

Dataset examples you can reference without redistributing

For restricted datasets like CO3D: include only official links and a script that checks for user-provided paths.
For open datasets: still prefer scripts + instructions, and avoid committing large archives.

Large files

If you must ship large non-dataset artifacts (e.g., small checkpoints, demo media):

use GitHub Releases for versioned artifacts, or
use Git LFS for limited large assets.

Training

python -m bayesfusionsdf.train   --config configs/train.yaml   --data_root ./data/<dataset_name>/processed   --output_dir ./outputs/exp01

Evaluation

python -m bayesfusionsdf.eval   --config configs/eval.yaml   --checkpoint ./outputs/exp01/checkpoints/latest.pt   --data_root ./data/<dataset_name>/processed   --output_dir ./outputs/exp01_eval

Reproducibility

Recommended practices:

keep fixed random seeds in configs,
log the exact command used for each run,
store environment details (Python version, package versions),
provide at least one minimal “smoke test” experiment that runs quickly.

Optional but recommended:

scripts/reproduce_main_results.sh
scripts/smoke_test.sh

Directory Structure

Suggested repository structure:

BayesFusionSDF/
  docs/                       # GitHub Pages site (custom domain)
    index.html
    CNAME
    paper.pdf                 # optional

  src/ or bayesfusionsdf/     # python package / core code
    __init__.py
    ...

  configs/
    train.yaml
    eval.yaml
    data/

  scripts/
    preprocess_*.py
    reproduce_*.sh
    smoke_test.sh

  assets/                     # small, permissive media only (figures, icons)
  tests/                      # optional unit/integration tests

  requirements.txt
  README.md
  LICENSE
  .gitignore
  DATASETS.md                 # recommended

Citation

If you use this repository in academic work, please cite:

@inproceedings{bayesfusionsdf2026,
  title     = {BayesFusionSDF},
  author    = {Mazumdar, Soumya and ...},
  booktitle = {Your Conference Name},
  year      = {2026}
}

License

Specify the license you intend to use (e.g., MIT, Apache-2.0). If your code depends on third-party libraries with additional requirements, list them in NOTICE or in this section.

Contact

Maintainer: Soumya Mazumdar
Website: https://soumyamazumdar.com

What files you should include (recommended checklist)

Must have

README.md (this file)
LICENSE
.gitignore
requirements.txt (or environment.yml)
docs/index.html and docs/CNAME (for the subdomain project page)

Strongly recommended

DATASETS.md (dataset setup + links + terms reminder)
scripts/preprocess_*.py (and optionally scripts/download_*.py)
configs/ for reproducible runs
scripts/smoke_test.sh (quick sanity run)

Optional

tests/
CHANGELOG.md
CITATION.cff (nice for GitHub “Cite this repository” button)

What you should NOT include

Any third-party dataset content (CO3D etc.)
Large raw data archives (.zip, .tar, .mp4 dumps) unless you own the rights and it’s intended
Secrets: API keys, tokens, private URLs (use .env.example instead)

This site is open source. Improve this page.