README.md

metadata

license: cc-by-nc-sa-4.0
language:
  - en
tags:
  - Audio
  - Video
  - Text
  - audio-visual
  - audio-visual reasoning
size_categories:
  - 1K<n<10K
task_categories:
  - question-answering
configs:
  - config_name: default
    data_files: savvy_bench.jsonl

SAVVY-Bench

This repository contains SAVVY-Bench, the first benchmark for dynamic 3D spatial reasoning in audio-visual environments, introduced in SAVVY: Spatial Awareness via Audio-Visual LLMs through Seeing and Hearing.

SAVVY-Bench Dataset

The benchmark dataset is also available on Hugging Face:

from datasets import load_dataset
dataset = load_dataset("ZijunCui/SAVVY-Bench")

This repository provides both the benchmark data and tools to process the underlying Aria Everyday Activities videos.

Setup Environment

Step 1: Create Conda Environment

# Create and activate conda environment (Python 3.10 recommended)
conda create -n savvy-bench python=3.10 -y
conda activate savvy-bench

# Install minimal dependencies for AEA processing
pip install requests tqdm numpy opencv-python imageio open3d matplotlib tyro pillow soundfile natsort

# Install Project Aria Tools (ESSENTIAL for VRS file processing) and VRS tools
pip install 'projectaria-tools[all]'
conda install -c conda-forge vrs

# Install PyTorch (required by EgoLifter rectification script)
# CPU version (sufficient for most users):
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu

# OR GPU version if you have NVIDIA GPU (optional):
# conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia

Step 2: Clone Repository with Submodules

# Clone the SAVVY-Bench repository
git clone --recursive https://github.com/ZijunCui/SAVVY-Bench.git
cd SAVVY-Bench

# The EgoLifter submodule will be automatically initialized with --recursive flag
# If you need to update submodules later, run:
# git submodule update --init --recursive

If you already cloned without submodules:

git submodule update --init --recursive

Step 3: Verify Installation

# Test essential dependencies
python -c "from projectaria_tools.core import mps; print('✓ ProjectAria tools')"
which vrs >/dev/null && echo "✓ VRS command available" || echo "✗ VRS command not found"
python -c "import torch; print('✓ PyTorch:', torch.__version__)"
python -c "import cv2, numpy, open3d; print('✓ OpenCV, NumPy, Open3D')"
python -c "import requests; print('✓ Download tools ready')"

Download and Process Aria Everyday Activities Videos

Step 1: Access the Dataset

1. Visit Aria Everyday Activities Dataset
2. Follow the instructions to access the dataset
3. Download the Aria Everyday Activities Dataset.json file and place it in the repository root

Step 2: Automatic Download and Undistortion

# Activate environment
conda activate savvy-bench

# Run the automated processing script
chmod +x aea.sh
./aea.sh

This script will:

• Download 52 AEA video sequences (.vrs format) and SLAM data with resume capability
• Extract RGB images and camera poses from VRS files using egolifter/scripts/process_project_aria_3dgs.py
• Remove fisheye distortion and rectify images using egolifter/scripts/rectify_aria.py
• Extract audio from VRS files
• Convert undistorted frames to MP4 videos
• Save all processed data in aea/aea_processed/

Step 3: Verify Processing

After completion, you should have:

• Raw data in aea/aea_data/ (52 scenes)
• Processed data in aea/aea_processed/ with the following structure:

aea/aea_processed/
├── loc1_script2_seq1_rec1/
│   ├── audio/
│   │   └── loc1_script2_seq1_rec1.wav  # Extract Audio from VRS
│   ├── video/
│   │   └── loc1_script2_seq1_rec1.mp4  # From undistorted frames
│   ├── images/                         # Undistorted frames
│   └── transforms.json                 # Camera poses for 3D reconstruction
├── loc1_script2_seq1_rec2/
│   └── ...
└── ...

Citation

@article{chen2025savvy,
    title={SAVVY: Spatial Awareness via Audio-Visual LLMs through Seeing and Hearing},
    author={Mingfei Chen and Zijun Cui and Xiulong Liu and Jinlin Xiang and Caleb Zheng and Jingyuan Li and Eli Shlizerman},
    year={2025},
    eprint={2506.05414},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}