mRobotit / DANet

DANet: Monocular Depth Distribution Alignment with Low Computation ICRA 2022

DANet: Monocular Depth Distribution Alignment with Low Computation

This is the implementation of the paper Monocular Depth Distribution Alignment with Low Computation, ICRA 2022, Fei Sheng, Feng Xue, Yicong Chang, Wenteng Liang and Anlong Ming.

Contents

1. Introduction
   
2. Usage
   
3. Results
   
4. Acknowledgements
   

Introduction

This repository provides the implementation of DANet in PyTorch.

Monocular depth estimation has been widely applied in many computer vision and robotics tasks. DANet is proposed to achieve a trade-off between speed and accuracy in this task.

Usage

Dependencies

• Python3.8
• PyTorch(1.10.0)
• Pytorch3d(0.6.1)
• Pandas(1.4.1)
• Scipy(1.6.2)
• skimage
• sklearn
• Wandb

Pre-processed Data

You can download the NYUD v2 and iBims-1 dataset from Google Drive or Baidu Netdisk (Code: dp53).
Then please unzip the data folder and replace the ./data folder.

Train

As an example, use the following command to train DANet on NYUDV2.

CUDA_VISIBLE_DEVICES="0,1" python train.py --logging (wandb logging during training)

Evaluation

Use the following command to evaluate the trained DANet on NYUDV2 test data.

CUDA_VISIBLE_DEVICES="0" python evaluate.py --loadckpt (the path of the loaded model)

Use the following command to evaluate the trained DANet on iBims-1 test data.

CUDA_VISIBLE_DEVICES="0" python evaluate_ibims1.py --loadckpt (the path of the loaded model)

Pretrained Model

You can download the pretrained model from Google Drive or Baidu Netdisk (Code: zjfy)

Results

Some examples

Citation

@inproceedings{Sheng2022DANet,
             title = {Monocular Depth Distribution Alignment with Low Computation},
             author = {Fei Sheng, Feng Xue, and Yicong Chang and Wenteng Liang
and Anlong Ming},
	     conference={IEEE International Conference on Robotics and Automation (ICRA)},
             year = {2022}   
} 

Acknowledgements

The source code of mini Vit in our method follows work. Our work is inspired by this work and part of codes.