This is the implementation of the paper Superpixel Cost Volume Excitation for Stereo Matching, PRCV 24, Shanglong Liu, Lin Qi, Junyu Dong, Wenxiang Gu, Liyi Xu[Arxiv].
Please contact Shanglong Liu ([email protected]) if you have any questions.
- NVIDIA RTX 3090
- python 3.7.13
- pytorch 1.12.0
To ensure connectivity during superpixel visualization, similar to SpixelFCN , we make use of the component connection method in SSN to enforce the connectivity in superpixels. The code has been included in /models/cython. To compile it:
cd models/cython/
python setup.py install --user
cd ../..
Training.
Before starting joint training, load the weights of the sub-network. The pre-trained weights should be available in the following path /pretrained/spixel_16/SpixelNet_bsd_ckpt.tar.
run the script ./scripts/sceneflow.sh to train on Scene Flow datsets. Please update DATAPATH in the bash file as your training data path.
Testing.
If only the training head is used, the sub-network can be omitted during the inference stage to maintain the same parameters and computational cost as the baseline network.
If you find our work useful in your research, please consider citing our paper:
@InProceedings{10.1007/978-981-97-8508-7_2,
author="Liu, Shanglong
and Qi, Lin
and Dong, Junyu
and Gu, Wenxiang
and Xu, Liyi",
title="Superpixel Cost Volume Excitation for Stereo Matching",
booktitle="Pattern Recognition and Computer Vision",
year="2025",
pages="18--31",
}