flybirdtian / nconv

A PyTorch implementation for our work "Confidence Propagation through CNNs for Guided Sparse Depth Regression"

NConv-CNN

This is the PyTorch implementation for our work:

1. Propagating Confidences through CNNs for Sparse Data Regression

2. Confidence Propagation through CNNs for Guided Sparse Depth Regression

If you use this code or compare against it, please cite our work:

@article{eldesokey2018propagating,
  title={Propagating Confidences through CNNs for Sparse Data Regression},
  author={Eldesokey, Abdelrahman and Felsberg, Michael and Khan, Fahad Shahbaz},
  journal={arXiv preprint arXiv:1805.11913},
  year={2018}
}

@article{eldesokey2018confidence,
  title={Confidence Propagation through CNNs for Guided Sparse Depth Regression},
  author={Eldesokey, Abdelrahman and Felsberg, Michael and Khan, Fahad Shahbaz},
  journal={arXiv preprint arXiv:1811.01791},
  year={2018}
}

Contents

0. Dependencies
1. Networks Description
2. Evaluation using Pretrained Weights
3. The NYU-Depth-v2 dataset
4. Contact

Dependecies

• opencv (To save output images)
• json (To read experiment parameters file)

Usage

python run-nconv-cnn.py -mode [MODE] -exp [EXP] -chkpt [CHKPT] -set [SET]

[MODE]: The mode could be either train or eval.

[EXP]: The name of the directory in 'workspace' which has the network file.

[CHKPT]: (optional) Continue traing from a specific epoch or evaluate using a specific epoch.

[SET]: (optional) The set to evaluate on. The possible options are val, selval or test.

Networks Description

Networks are located in "workspace" directory. Each network file is stored in its own directory and associated with params.json which has the training parameters for the network.

Four netwokrs are available:

1. exp_unguided_disparity: Unguided depth completion network trained on disparity (Deonted as NConv-HMS in paper [1])
2. exp_unguided_depth: Unguided depth completion network trained on depth.
3. exp_guided_nconv_cnn_l1: Guided depth completion network trained on depth (Denoted as MS-Net[LF] or NConv-CNN-L1 in paper [2])
4. exp_guided_nconv_cnn_l2: Guided depth completion network trained on depth (Denoted as NConv-CNN-L2 in paper [2])
5. exp_guided_enc_dec: Guided depth completion network trained on depth (Denoted as Enc-Dec[EF] in paper [2])

Evaluation using Pretrained Weights

You can evaluate any of the networks using the pretrained-weights by calling

python run-nconv-cnn.py -mode eval -exp <exp_name>

The NYU-Depth-v2 dataset

The implemntation for the NYU-Depth-v2 dataset can be found at: https://github.com/abdo-eldesokey/nconv-nyu

Contact

Abdelrahman Eldesokey

E-mail: [email protected]