XiangMochu / MDEUncertainty

Measuring and Modeling Uncertainty Degree for Monocular Depth Estimation

[arXiv]

Introduction

This is a PyTorch implementation of Measuring and Modeling Uncertainty Degree for Monocular Depth Estimation.

Training

Data preparation

Please change the path to the NYU V2 zip file in utils/options.py (check out DenseDepth to download the zipfile):

DEFAULTS = {'nyu_data_path':
                {'your-pc-name-here': 'your-data-path-here'},
            ...}

Please chenge the path to the KITTI dataset in utils/kitti_data.py:

class KittiDefaultArg(object):
    def __init__(self, opts):
        self.dataset = 'kitti'
        self.filenames_file = 'utils/kitti/kitti_eigen_train_files_with_gt_new.txt'
        self.filenames_file_eval = 'utils/kitti/kitti_eigen_test_files_with_gt.txt'
        self.data_path = 'MODIFY_ME/kitti_data'
        self.gt_path = 'MODIFY_ME/data_depth_annotated'

        self.data_path_eval = 'MODIFY_ME/kitti_data'
        self.gt_path_eval = 'MODIFY_ME/data_depth_annotated'

Requirements

pip install -r requirements.txt

Training

There are a few options to train an MDE network, you can choose between:

dataset: nyu, kitti

encoder: resnet, densenet, swin, vit, efficientnet

decoder: simple, bts

reg_mode: direct, lin_cls

reg_supervision: regression_l1, none

prob_supervision: soft_label, none

uncert_supervision: error_uncertainty_ranking, error_uncertainty_ranking_noclamp, error_uncertainty_l1, none

You can choose between these options by adding extra arguments, for example:

python train.py \
  --dataset nyu \
  --encoder swin \
  --reg_mode lin_cls \
  --reg_supervision regression_l1 \
  --prob_supervision soft_label \
  --uncert_supervision error_uncertainty_ranking

Evaluation

To evaluate the accuracy and uncertainty degree of a trained model, run eval.py and add the path to the model checkpoint folder as an extra artument.

python eval.py Res_Sim_Lin_L-1_Non_Eur_kitti_2023_02_28-22:42:51

Citation

@article{xiang2023measuring,
  title={Measuring and Modeling Uncertainty Degree for Monocular Depth Estimation},
  author={Xiang, Mochu and Zhang, Jing and Barnes, Nick and Dai, Yuchao},
  journal={arXiv preprint arXiv:2307.09929},
  year={2023}
}