ADStereo: Efficient Stereo Matching with Adaptive Downsampling and Disparity Alignment TIP 2025

This paper presents two sampling strategies: the Adaptive Downsampling Module (ADM) and the Disparity Alignment Module (DAM), to prioritize real-time inference while ensuring accuracy. The ADM leverages local features to learn adaptive weights, enabling more effective downsampling while preserving crucial structure information. On the other hand, the DAM employs a learnable interpolation strategy to predict transformation offsets of pixels, thereby mitigating the spatial misalignment issue. Building upon these modules, we introduce ADStereo, a real-time yet accurate network that achieves highly competitive performance on multiple public benchmarks.

Demo on KITTI raw data

The pretrained KITTI model is loaded from './fined/KITTI/' datafolders.

ADStereo.mp4

Run demo_video.py to perform stereo matching on the raw Kitti sequence. Here is an example result on our system with RTX a5000ada on Ubuntu 20.04

Adaptive Downsampling Module & Disparity Alignment Module

Overview

New Added

We introduce a more lightweight model called ADStereo_fast (highly competetive performance & faster speed), also included in this repo.

Quantative Results

Model Zoo

All pretrained models are available in the Google Driver:ADStereo and Google Driver:ADStereo_fast

We assume the downloaded weights are located under the ./trained directory.

Otherwise, you may need to change the corresponding paths in the scripts.

Environment

Python 3.9
Pytorch 2.4.0

Create a virtual environment and activate it.

conda create -n ADStereo python=3.9
conda activate ADStereo

Dependencies

conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch -c nvidia
pip install opencv-python
pip install scikit-image
pip install tensorboard
pip install matplotlib 
pip install tqdm
pip install chardet
pip install imageio
pip install thop
pip install timm==0.5.4

1. Prepare training data

To evaluate/train ADStereo, you will need to download the required datasets.

SceneFlow

KITTI

Middlebury

ETH3D

By default datasets.py will search for the datasets in these locations.

DATA
├── KITTI
│   ├── kitti_2012
│   │   └── training
        └── testing
│   ├── kitti_2015
│   │   └── training
        └── testing
└── SceneFlow
    ├── Driving
    │   ├── disparity
    │   └── frames_finalpass
    ├── FlyingThings3D
    │   ├── disparity
    │   └── frames_finalpass
    └── Monkaa
        ├── disparity
        └── frames_finalpass
└── Middlebury
    ├── trainingH
    ├── trainingH_GT
└── ETH3D
    ├── two_view_training
    ├── two_view_training_gt

2. Train on SceneFlow

Run main.py to train on the SceneFlow dataset. Please update datapath in main.py as your training data path.

3. Finetune & Inference

Run finetune.py to finetune on the different real-world datasets, such as KITTI 2012, KITTI 2015, and ETH3D. Please update datapath in finetune.py as your training data path.

4. Evaluate FLOPs

Run counts_op.py to validate FLOPs consumption.

5. Results

To generate prediction results on the test set of the KITTI dataset, you can run evaluate_kitti.py. The inference time can be printed once you run evaluate_kitti.py. And the inference results on the KITTI dataset can be directly submitted to the online evaluation server for benchmarking.

Citation

If you find our work useful in your research, please consider citing our paper:

@article{wang2025ad,
  author={Wang, Yun and Li, Kunhong and Wang, Longguang and Hu, Junjie and Wu, Dapeng Oliver and Guo, Yulan},
  journal={IEEE Transactions on Image Processing}, 
  title={ADStereo: Efficient Stereo Matching with Adaptive Downsampling and Disparity Alignment}, 
  journal={IEEE Transactions on Image Processing},
  year={2025},
  publisher={IEEE}
}

Acknowledgements

This project is based on GwcNet, IGEV-Stereo, and CoEx. We thank the original authors for their excellent works.