pcs-stereo

This repository is code for our paper "Proxy-supervised Cross-spectral Stereo Matching"

The pipeline of our method is as follows:

Enviroment

Our experiments were conducted in the following environments:

• Nvidia GForce 3090 * 1

• Ubuntu 18.04

• Python 3.8

• Pytorch

For detailed environment configuration, please refer to requirements.txt

special dependency installation

pip install 'git+https://github.com/saadnaeem-dev/pytorch-linear-warmup-cosine-annealing-warm-restarts-weight-decay'

Data preparation

We use Pittsburgh cross-spectrial stereo dataset, please refer to DMC for downloading.

To generate dense pseudo-labels， please first follow the steps in the paper to generate initial labels using Metric3D and CREStereo, then refer to our code in pseudo_label_generation folder.

more qualitative results of our pseudo-label generation method are shown in imgs/results.pdf.

Pretrained Model

Download pretrained models

Performance (RMSE, lower is better):

Model	Common	Light	Glass	Glossy	Vegetation	Skin	Clothing	Bag	Mean
PSMNet*	0.45	0.79	0.83	0.99	0.65	0.83	0.83	0.59	0.74
IGEVStereo*	0.42	0.46	0.82	0.95	0.59	0.58	0.44	0.50	0.60

'*' denotes the conventional stereo-matching network trained using our method.

Train and evaluation

Please refer to run.ipynb for details.