XinyuHou97 / ThermalSfMLearner-MS

Official implementation of the paper "Self-supervised Depth and Ego-motion Estimation for Monocular Thermal Video using Multi-spectral Consistency Loss"

Self-supervised Depth and Ego-motion Estimation for Monocular Thermal Video using Multi-spectral Consistency Loss

This github implements the system described in the paper:

Self-supervised Depth and Ego-motion Estimation for Monocular Thermal Video using Multi-spectral Consistency Loss

Ukcheol Shin, Kyunghyun Lee, Seokju Lee, In So Kweon

Robotics and Automation Letter 2022 & ICRA 2022

[PDF] [Project webpage] [Full paper] [Youtube]

Depth estimation results on ViViD dataset

Prerequisite

This codebase was developed and tested with python 3.7, Pytorch 1.5.1, and CUDA 10.2 on Ubuntu 16.04.

conda env create --file environment.yml

Datasets

See "scripts/run_prepare_vivid_data.sh".

For ViViD Raw dataset, download the dataset provided on the official website.

For our post-processed dataset and pre-trained models, you can download after fill out a simple survey.

We will send you an e-mail with a download link.

After download post-processed dataset, generate training/testing dataset

sh scripts/run_prepare_vivid_data.sh

Training

The "scripts" folder provides several examples for training and testing.

You can train the depth and pose model on vivid dataset by running

sh scripts/train_vivid_resnet18_indoor.sh
sh scripts/train_vivid_resnet18_outdoor.sh

Then you can start a tensorboard session in this folder by

tensorboard --logdir=checkpoints/

and visualize the training progress by opening https://localhost:6006 on your browser.

Evaluation

You can evaluate depth and pose by running

sh scripts/test_vivid_indoor.sh
sh scripts/test_vivid_outdoor.sh

and visualize depth by running

sh scripts/run_inference.sh

You can comprehensively see the overall results by running

sh scripts/display_result.sh

Depth Results

Indoor

Models	Abs Rel	Sq Rel	RMSE	RMSE(log)	Acc.1	Acc.2	Acc.3
Ours(T)	0.231	0.215	0.730	0.266	0.616	0.912	0.990
Ours(MS)	0.163	0.123	0.553	0.204	0.771	0.970	0.995

Outdoor

Models	Abs Rel	Sq Rel	RMSE	RMSE(log)	Acc.1	Acc.2	Acc.3
Ours(T)	0.157	1.179	5.802	0.211	0.750	0.948	0.985
Ours(MS)	0.146	0.873	4.697	0.184	0.801	0.973	0.993

Pose Estimation Results

Indoor-static-dark

Metric	ATE	RE
Ours(T)	0.0063	0.0092
Ours(MS)	0.0057	0.0089

Outdoor-night1

Metric	ATE	RE
Ours(T)	0.0571	0.0280
Ours(MS)	0.0562	0.0287

Citation

Please cite the following paper if you use our work, parts of this code, and pre-processed dataset in your research.

@article{shin2021self,
  title={Self-Supervised Depth and Ego-Motion Estimation for Monocular Thermal Video Using Multi-Spectral Consistency Loss},
  author={Shin, Ukcheol and Lee, Kyunghyun and Lee, Seokju and Kweon, In So},
  journal={IEEE Robotics and Automation Letters},
  volume={7},
  number={2},
  pages={1103--1110},
  year={2021},
  publisher={IEEE}
}

Related projects

• SfMLearner-Pytorch (CVPR 2017)
• SC-SfMLearner-Pytorch (NeurIPS 2019)