DMS: Diffusion-Based Multi-Baseline Stereo Generation for Improving Self-Supervised Depth Estimation

ICCV 2025 – AIM Workshop (Oral)
DMS leverages diffusion models to synthesize epipolar-aligned multi-baseline views (left-shifted, right-shifted, and intermediate) that explicitly complete occluded and out-of-frame regions—boosting self-supervised depth learning without extra labels.

---

TL;DR

• 🎯 Goal: Improve self-supervised stereo/mono depth in ill-posed regions.
• 🧠 Idea: Fine-tune a diffusion UNet to generate novel views guided by simple directional prompts (“to left / to right / middle”).
• 🚀 Win: Adds valid correspondences where photometric supervision was previously missing.

---

Table of Contents

• Installation
• Data Preparation
• Pretrained Models
• Train the DMB Diffusion Model
• Inference: Multi-Baseline Images
• Citation
• License
• Contact

---

Installation

pip install -r requirements.txt

---

Data Preparation

Please download the following datasets:

• SceneFlow
• KITTI Raw
• KITTI 2012 & 2015
• MPI-Sintel

Set your dataset roots in the corresponding config or script files as needed.

---

Pretrained Models (Google Drive)

• SceneFlow-DMB
• KITTI-Raw-DMB
• KITTI-2012-DMB
• KITTI-2015-DMB
• MPI-Sintel-DMB

---

Train the DMB Diffusion Model

SceneFlow

cd scripts/SF/train
sh train_unet.sh

KITTI Raw

cd scripts/KITTI/train
sh train_kitti_raw.sh

KITTI 2015

cd scripts/KITTI/train
sh train_kitti15.sh

KITTI 2012

cd scripts/KITTI/train
sh train_kitti12.sh

MPI-Sintel

cd scripts/MPI/train
sh train_unet.sh

---

Inference: Multi-Baseline Images

SceneFlow

# left->right and right->left
cd scripts/SF/evaluation
sh evaluation.sh

# left-left and right-right
cd scripts/SF/evaluation
sh get_additional_view.sh

# middle-state views
cd scripts/SF/evaluation
sh get_middle_view.sh

KITTI Raw

# left->right and right->left
cd scripts/KITTI/kitti_raw_evaluations
sh eval_unet.sh

# left-left and right-right
cd scripts/KITTI/kitti_raw_evaluations
sh unet_generated_new_view.sh

# middle-state views
cd scripts/KITTI/kitti_raw_evaluations
sh unet_generate_med_view.sh

KITTI 2015

# left->right and right->left
cd scripts/KITTI/kitti2015_evaluations
sh unet_eval.sh

# left-left and right-right
cd scripts/KITTI/kitti2015_evaluations
sh get_additional_view.sh

# middle-state views
cd scripts/KITTI/kitti2015_evaluations
sh get_middle_view.sh

KITTI 2012

# left->right and right->left
cd scripts/KITTI/kitti2012_evaluations
sh unet_eval.sh

# left-left and right-right
cd scripts/KITTI/kitti2012_evaluations
sh get_additional_view.sh

# middle-state views
cd scripts/KITTI/kitti2012_evaluations
sh get_middle_view.sh

MPI-Sintel

# left/right + left-left/right-right
cd scripts/MPI/evaluations
sh eval_unet.sh

# middle-state views
cd scripts/MPI/evaluations
sh unet_generate_med_view.sh

---

Citation

If you find this work helpful, please cite:

@inproceedings{liu2025dms,
  title     = {DMS: Diffusion-Based Multi-Baseline Stereo Generation for Improving Self-Supervised Depth Estimation},
  author    = {Liu, Zihua and {co-authors}},
  booktitle = {ICCV Workshops (AIM)},
  year      = {2025},
}

---

License

This project is released under the MIT License. See LICENSE for details.

---

Contact

For questions, please open an issue or contact the authors via the project page.