MonoProb: Self-Supervised Monocular Depth Estimation with Interpretable Uncertainty

This repository contains the official implementation of MonoProb: Self-Supervised Monocular Depth Estimation with Interpretable Uncertainty that has been accepted to the IEEE Winter Conference on Applications of Computer Vision (WACV) 2024.

Requirements

See requirements.txt.

Data

Download KITTI raw dataset and accurate ground truth maps:

sh scripts/prepare_kitti_data.sh /path/to/kitti_data

Training

MonoProb M without self-distillation:

python train.py \
--model_name model_name \
--data_path /path/to/kitti_data \
--log_dir /path/to/save/checkpoints \
--uncertainty \
--distribution normal \
--sample_size 9 \
--uncert_act sigmoid \
--uncert_as_a_fraction_of_depth \

MonoProb M with self-distillation:

python train.py \
--model_name model_name \
--data_path /path/to/kitti_data \
--log_dir /path/to/save/checkpoints \
--uncertainty \
--distribution normal \
--self \
--load_weights_folder /path/of/the/teacher's/checkpoints \
--uncert_act no \
--models_to_load encoder depth

NB: Use the last checkpoints returned by MonoProb M without self-distillation as teacher's checkpoints.

Options for the other training paradigms:

• Training from stereo images only (S): --use_stereo --frame_ids 0 --split eigen_full
• Training from monocular videos and stereo images only (MS): --use_stereo

Evaluation

MonoProb checkpoints are available here. Download all checkpoints with:

sh scripts/download_checkpoints.sh

Evaluation scripts:

sh scripts/eval_M.sh /path/to/kitti_data
sh scripts/eval_S.sh /path/to/kitti_data
sh scripts/eval_MS.sh /path/to/kitti_data

Citation

@misc{marsal2023monoprob,
      title={MonoProb: Self-Supervised Monocular Depth Estimation with Interpretable Uncertainty}, 
      author={Rémi Marsal and Florian Chabot and Angelique Loesch and William Grolleau and Hichem Sahbi},
      year={2023},
}

Acknowledgements

We thank the authors of Monodepth2 and of Mono-uncertainty for their great work and for sharing their code.