Fusion-then-Distillation: Toward Cross-modal Positive Distillation for Domain Adaptive 3D Semantic Segmentation [TCSVT 2025]

🔥 This is the official PyTorch implementation of our work. 🔥

Overview

News

• 2025-02 📢 This work has been accepted by IEEE Transactions on Circuits and Systems for Video Technology (TCSVT).
• 2024-10 📢 Our new work, UniDSeg, has been accepted by NeurIPS 2024. Code
• 2024-09 🚀 We release the code for FtD-PlusPlus. It's an extended version of BFtD.

Getting Started

Installation

The implementation runs on

• Python 3.8
• Torch 1.9.0
• Torchvision 0.10.0
• CUDA 11.1
• SparseConvNet
• nuscenes-devkit

Preparation

As 3D network we use SparseConvNet. We advise to create a new conda environment for installation. PyTorch and CUDA can be installed, and SparseConvNet installed/compiled as follows:

pip install --upgrade git+https://github.com/facebookresearch/SparseConvNet.git

Dataset (Refer to xMUDA_journal)

You need to perform preprocessing to generate the data for FtD-PlusPlus.

NuScenes:

• Please download the Full dataset (v1.0) from the NuScenes website and extract it.
• Please edit the script xmuda/data/nuscenes/preprocess.py as follows and then run it.
  • root_dir should point to the root directory of the NuScenes dataset.
  • out_dir should point to the desired output directory to store the pickle files.

A2D2:

• Please download the Semantic Segmentation dataset and Sensor Configuration from the Audi website or directly use wget and the following links, then extract.
• Please edit the script xmuda/data/a2d2/preprocess.py as follows and then run it.
  • root_dir should point to the root directory of the A2D2 dataset.
  • out_dir should point to the desired output directory to store the undistorted images and pickle files. It should be set differently than the root_dir to prevent overwriting of images.

SemanticKITTI:

• Please download the files from the SemanticKITTI website and additionally the color data from the Kitti Odometry website. Extract everything into the same folder. Similar to NuScenes preprocessing, we save all points that project into the front camera image as well as the segmentation labels to a pickle file.
• Please edit the script xmuda/data/semantic_kitti/preprocess.py as follows and then run it.
  • root_dir should point to the root directory of the SemanticKITTI dataset.
  • out_dir should point to the desired output directory to store the pickle files.

VirtualKITTI:

• Clone the repo from VirtualKITTI website.
• Download raw data and extract with following script.

  cd vkitti3D-dataset/tools
  mkdir path/to/virtual_kitti
  bash download_raw_vkitti.sh path/to/virtual_kitti
  

• Generate point clouds (npy files).

  cd vkitti3D-dataset/tools
  for i in 0001 0002 0006 0018 0020; do python create_npy.py --root_path path/to/virtual_kitti --out_path path/to/virtual_kitti/vkitti_npy --sequence $i; done
  

• Similar to NuScenes preprocessing, we save all points and segmentation labels to a pickle file.
• Please edit the script xmuda/data/virtual_kitti/preprocess.py as follows and then run it.
  • root_dir should point to the root directory of the VirtualKITTI dataset.
  • out_dir should point to the desired output directory to store the pickle files.

Usage

Training UDA

cd <root dir of this repo>
# day to night
CUDA_VISIBLE_DEVICES=0 python -W ignore xmuda/train_dn_bftd_uni.py --cfg=configs/nuscenes_lidarseg/day_night/uda/bftd.yaml
# usa to singapore
CUDA_VISIBLE_DEVICES=0 python -W ignore xmuda/train_us_bftd_uni.py --cfg=configs/nuscenes_lidarseg/usa_singapore/uda/bftd.yaml
# virtualkitti to semantickitti
CUDA_VISIBLE_DEVICES=0 python -W ignore xmuda/train_vs_bftd_uni.py --cfg=configs/virtual_kitti_semantic_kitti/uda/bftd.yaml
# a2d2 to semantickitti
CUDA_VISIBLE_DEVICES=0 python -W ignore xmuda/train_as_bftd_uni.py --cfg=configs/a2d2_semantic_kitti/uda/bftd.yaml

Training SSDA

cd <root dir of this repo>
# usa to singapore
CUDA_VISIBLE_DEVICES=0 python -W ignore xmuda/train_us_bftd_uni_ssda.py --cfg=configs/nuscenes_lidarseg/usa_singapore/ssda/bftd.yaml
# a2d2 to semantickitti
CUDA_VISIBLE_DEVICES=0 python -W ignore xmuda/train_as_bftd_uni_ssda.py --cfg=configs/a2d2_semantic_kitti/ssda/bftd.yaml

You can change the path OUTPUT_DIR in the config file bftd.yaml and bftd_pl.yaml.

Testing

You can run the testing with:

cd <root dir of this repo>
CUDA_VISIBLE_DEVICES=0 python -W ignore xmuda/test.py \
    --cfg=./configs/nuscenes/day_night/uda/bftd.yaml \
    --ckpt2d=./output/day_night/FtD-PlusPlus/best_model_2d.pth \
    --ckpt3d=./output/day_night/FtD-PlusPlus/best_model_3d.pth \
    --fusion=./output/day_night/FtD-PlusPlus/best_model_fusion.pth

Re-training with pseudo-labels

After having trained the FtD-PlusPlus model, generate the offline pseudo-labels as follows:

cd <root dir of this repo>
CUDA_VISIBLE_DEVICES=0 python -W ignore xmuda/test.py \
    --cfg=configs/nuscenes/day_night/uda/bftd_pl.yaml \
    --ckpt2d=./output/day_night/FtD-PlusPlus/model_2d_100000.pth \
    --ckpt3d=./output/day_night/FtD-PlusPlus/model_3d_100000.pth \
    --fusion=./output/day_night/FtD-PlusPlus/model_fusion_100000.pth \
    --pselab \
    DATASET_TARGET.TEST "('train_night',)"

Note that we use the last model at 100,000 steps to exclude supervision from the validation set by picking the best weights. The pseudo labels and maximum probabilities are saved as .npy file.

Please edit the pselab_paths in the config file, e.g. configs/nuscenes/day_night/bftd_pl.yaml, to match your path of the generated pseudo-labels. Then start the training. The pseudo-label refinement (discard less confident pseudo-labels) is done when the dataloader is initialized.

cd <root dir of this repo>
# UDA
CUDA_VISIBLE_DEVICES=0 python -W ignore xmuda/train_dn_bftd_uni.py --cfg=configs/nuscenes_lidarseg/day_night/uda/bftd_pl.yaml
CUDA_VISIBLE_DEVICES=0 python -W ignore xmuda/train_us_bftd_uni.py --cfg=configs/nuscenes_lidarseg/usa_singapore/uda/bftd_pl.yaml
CUDA_VISIBLE_DEVICES=0 python -W ignore xmuda/train_vs_bftd_uni.py --cfg=configs/virtual_kitti_semantic_kitti/uda/bftd_pl.yaml
CUDA_VISIBLE_DEVICES=0 python -W ignore xmuda/train_as_bftd_uni.py --cfg=configs/a2d2_semantic_kitti/uda/bftd_pl.yaml
# SSDA
CUDA_VISIBLE_DEVICES=0 python -W ignore xmuda/train_us_bftd_uni_ssda.py --cfg=configs/nuscenes_lidarseg/usa_singapore/ssda/bftd_pl.yaml
CUDA_VISIBLE_DEVICES=0 python -W ignore xmuda/train_as_bftd_uni_ssda.py --cfg=configs/a2d2_semantic_kitti/ssda/bftd_pl.yaml

Experimental Results

Acknowledgements

Code is built based on xMUDA_journal and BFtD.

Citation

If you find this project useful, please consider citing:

@article{DBLP:journals/corr/abs-2410-19446,
  author       = {Yao Wu and
                  Mingwei Xing and
                  Yachao Zhang and
                  Yuan Xie and
                  Yanyun Qu},
  title        = {Fusion-then-Distillation: Toward Cross-modal Positive Distillation
                  for Domain Adaptive 3D Semantic Segmentation},
  journal      = {CoRR},
  volume       = {abs/2410.19446},
  year         = {2024},
  url          = {https://doi.org/10.48550/arXiv.2410.19446},
  doi          = {10.48550/ARXIV.2410.19446},
  eprinttype   = {arXiv},
  eprint       = {2410.19446},
  timestamp    = {Thu, 28 Nov 2024 21:32:42 +0100},
  biburl       = {https://dblp.org/rec/journals/corr/abs-2410-19446.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}