🛹 RollingDepth: Video Depth without Video Models

CVPR 2025

This repository represents the official implementation of the paper titled "Video Depth without Video Models".

Bingxin Ke¹, Dominik Narnhofer¹, Shengyu Huang¹, Lei Ke², Torben Peters¹, Katerina Fragkiadaki², Anton Obukhov¹, Konrad Schindler¹

¹ETH Zurich, ²Carnegie Mellon University

📢 News

2025-02-26: Paper is accepted to CVPR 2025. <br> 2024-12-02: Paper is on arXiv.<br> 2024-11-28: Inference code is released.<br>

🛠️ Setup

The inference code was tested on: Debian 12, Python 3.12.7 (venv), CUDA 12.4, GeForce RTX 3090

📦 Repository

git clone https://github.com/prs-eth/RollingDepth.git
cd RollingDepth

🐍 Python environment

Create python environment:

# with venv
python -m venv venv/rollingdepth
source venv/rollingdepth/bin/activate

# or with conda
conda create --name rollingdepth python=3.12
conda activate rollingdepth

💻 Dependencies

Install dependicies:

pip install -r requirements.txt
bash script/install_diffusers_dev.sh  # Install modified diffusers with cross-frame self-attention

We use pyav for video I/O, which relies on ffmpeg (tested with version 5.1.6-0+deb12u1).

To see the modification in diffusers, search for comments "Modified in RollingDepth".

🏃 Test on your videos

All scripts are designed to run from the project root directory.

📷 Prepare input videos

1. Use sample videos:

   bash script/download_sample_data.sh
   

   These example videos are to be used only as debug/demo input together with the code and should not be distributed outside of the repo.

2. Or place your videos in a directory, for example, under data/samples.

🚀 Run with presets

python run_video.py \
    -i data/samples \
    -o output/samples_fast \
    -p fast \
    --verbose

• -p or --preset: preset options
  • fast for fast inference, with dilations [1, 25] (flexible), fp16, without refinement, at max. resolution 768.
  • fast1024 for fast inference at resolution 1024
  • full for better details, with dilations [1, 10, 25] (flexible), fp16, with 10 refinement steps, at max. resolution 1024.
  • paper for reproducing paper numbers, with (fixed) dilations [1, 10, 25], fp32, with 10 refinement steps, at max. resolution 768.
• -i or --input-video: path to input data, can be a single video file, a text file with video paths, or a directory of videos.
• -o or --output-dir: output directory.

Passing these inference arguments will overwrite the preset settings:

• --res or --processing-resolution: the maximum resolution (in pixels) at which image processing will be performed. If set to 0, processes at the original input image resolution.
• --refine-step: number of refinement iterations to improve accuracy and details. Set to 0 to disable refinement.
• --snip-len or --snippet-lengths: number of frames to analyze in each snippet.
• -d or --dilations: spacing between frames for temporal analysis, could have multiple values e.g. -d 1 10 25.

Clip sub-sequence to be processed:

• --from or --start-frame: the starting frame index for processing, default to 0.
• --frames or --frame-count: number of frames to process after the starting frame. Set to 0 (default) to process until the end of the video.

Output settings

• --fps or --output-fps: frame rate (FPS) for the output video. Set to 0 (default) to match the input video's frame rate.
• --restore-res or --restore-resolution: whether to restore the output to the original input resolution after processing, Default: False.
• --save-sbs or --save-side-by-side: whether to save side-by-side videos of RGB and colored depth. Default: True.
• --save-npy: whether to save depth maps as .npy files. Default: True.
• --save-snippets: whether to save initial snippets. Default: False

Other argumenets

• Please run python run_video.py --help to get details for other arguments.
• For low GPU memory footage: pass --max-vae-bs 1 --unload-snippet true and use a smaller resolution, e.g. --res 512

⬇ Checkpoint cache

By default, the checkpoint is stored in the Hugging Face cache. The HF_HOME environment variable defines its location and can be overridden, e.g.:

export HF_HOME=$(pwd)/cache

Alternatively, use the following script to download the checkpoint weights locally and specify checkpoint path by -c checkpoint/rollingdepth-v1-0

bash script/download_weight.sh

🦿 Evaluation on test datasets

Coming soon

🎓 Citation

@InProceedings{ke2024rollingdepth,
    title={Video Depth without Video Models}, 
    author={Bingxin Ke and Dominik Narnhofer and Shengyu Huang and Lei Ke and Torben Peters and Katerina Fragkiadaki and Anton Obukhov and Konrad Schindler},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year={2025}
}

🙏 Acknowledgments

We thank Yue Pan, Shuchang Liu, Nando Metzger, and Nikolai Kalischek for fruitful discussions.

We are grateful to redmond.ai (robin@redmond.ai) for providing GPU resources.

🎫 License

This code of this work is licensed under the Apache License, Version 2.0 (as defined in the LICENSE).

The model is licensed under RAIL++-M License (as defined in the LICENSE-MODEL)

By downloading and using the code and model you agree to the terms in LICENSE and LICENSE-MODEL respectively.