Run Docker to build environment usage: bash run_docker.sh donghe
IROS, 2023
Yunze Man
·
Liang-Yan Gui
·
Yu-Xiong Wang
·
This repository contains a pytorch implementation for the paper: DualCross: Cross-Modality Cross-Domain Adaptation for Monocular BEV Perception. This work results in the first open analysis of cross-domain cross-sensor perception and adaptation for monocular 3D tasks in the wild.
Clone this repository:
git clone https://github.com/YunzeMan/DualCross
cd DualCrossInstall requirements in a virtual environment (Tested on anaconda/miniconda environment):
conda create --name dualcross --file requirements.txtDownload nuscenes data from https://www.nuscenes.org/. Install dependencies.
pip install nuscenes-devkit tensorboardX efficientnet_pytorch==0.7.0
The code is tested on Python 3.6.10 with Pytorch 1.5.0+cu102 and 1.7.0+cu110.
To train the LiDAR-Teacher model first
python main.py train --logdir=./runs-lteacher --dataroot=/directory/to/dataset --bsz=8 --gpuid=0 --domain_gap=True --source='day' --target='night' --parser_name='lidarinputdata' --up_scale=4 --color_jitter=True --rand_resize=False --teacher_student=True --adv_training=FalseTo train the Image-Student model
python main.py train --logdir=./runs-istudent --dataroot=/directory/to/dataset --bsz=4 --gpuid=0 --domain_gap=True --source='day' --target='night' --parser_name='imglidardata' --up_scale=4 --color_jitter=True --rand_resize=False --adv_training=True --align_place='midfinal' --domainloss_lambda=0.1 --middomain_weight=0.1 --teacher_student=True --train_student=True --modelf_teacher=/directory/to/trained/teacher/model --teacher_lambda=1.0 --use_gt=True --gt_lambda=0.1 --use_depth=True --depth_lambda=0.05python main.py viz_model_preds --domain_gap=True --source='day' --target='night' --gpuid=1 --nsweeps=1 --up_scale=4 --bsz=1 --version='trainval' --modelf=/directory/to/trained/model --parser_name='vizdata' --viz_gt=False --viz_depth=True --mask_2d=True --sep_dpft=False --strict_import=True --adv_training=True --align_midfin=Truepython main.py make_video --vis_folder='./visualization' --vtype='tags' --start_frame=0 --end_frame=150A comparison between Cross-domain Adaptation (Left) and DualCross (Right) on day -> night scenario.
If you find our work useful in your research, please cite:
@inproceedings{man2023dualcross,
title={{DualCross: Cross-Modality Cross-Domain Adaptation for Monocular BEV Perception}},
author={Man, Yunze and Gui, Liang-Yan and Wang, Yu-Xiong},
booktitle={2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
year={2023}
}This repo is built based on the fantastic work Lift-Splat-Shoot by Jonah Philion and Sanja Fidler.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent