Reliable Student: Addressing Noise in Semi-Supervised 3D Object Detection

This is the official implementation of the paper "Reliable Student: Addressing Noise in Semi-Supervised 3D Object Detection" [arXiv].

Results

Evaluation results on the KITTI dataset.

Labeled Percent		1%			2%
Method	Car	Pedestrian	Cyclist	Car	Pedestrian	Cyclist
PV-RCNN	74.1	31.7	28.8	76.8	40.4	42.3
3DIoUMatch (Baseline)	76.4	35.7	36.0	78.9	47.0	53.3
ReliableStudent	77.0	41.9	36.4	79.5	53.0	59.0

Installation

# Create a conda environment
conda create -n reliable_student python=3.8
conda activate reliable_student

# Install PyTorch 1.7.1 with CUDA 11.0
conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=11.0 -c pytorch

# Install spconv 1.2.1
git clone -b v1.2.1 https://github.com/traveller59/spconv.git --recursive
cd spconv
python setup.py bdist_wheel
python -m pip install ./dist/spconv-1.2.1-cp38-cp38-linux_x86_64.whl
cd ..

# Clone the repo, install requirements and compile
git clone https://github.com/fnozarian/ReliableStudent.git
cd ReliableStudent
python -m pip install -r requirements.txt
python setup.py develop

Getting Started

Data Preparation

Please refer to Data Preparation to generate the data splits or use the baseline data we provide.

Pre-training

# Single GPU (e.g., A100-80GB)
python -u pretrain.py \
    --cfg_file ./cfgs/kitti_models/pv_rcnn.yaml \
    --batch_size 8 \
    --split train_0.01_1 \
    --extra_tag train_0.01_1 \
    --ckpt_save_interval 4 \
    --repeat 10 \
    --dbinfos kitti_dbinfos_train_0.01_1_37.pkl

# Multi-GPU with Slurm (e.g., A100-40GB x2)
srun -p A100-40GB --job-name=pretrain_0.01_1 --ntasks=2 \
    --gpus-per-task=1 --cpus-per-task=3 \
    --mem=60GB --kill-on-bad-exit=1 python -u pretrain.py \
    --cfg_file ./cfgs/kitti_models/pv_rcnn.yaml --batch_size 8 \
    --launcher slurm --split train_0.01_1 --extra_tag train_0.01_1 \
    --ckpt_save_interval 4 --repeat 10 \
    --dbinfos kitti_dbinfos_train_0.01_1_37.pkl

Training

# Single GPU (e.g., A100-80GB)
python -u train.py \
    --cfg_file ./cfgs/kitti_models/pv_rcnn_ssl_60.yaml \
    --batch_size 8 \
    --launcher slurm \
    --split train_0.01_1 \
    --extra_tag train_0.01_1 \
    --ckpt_save_interval 2 \
    --pretrained_model "../output/cfgs/kitti_models/pv_rcnn/split_0.01_1/ckpt/checkpoint_epoch_80.pth" \
    --repeat 5 \
    --dbinfos kitti_dbinfos_train_0.01_1_37.pkl
    
# Multi-GPU with Slurm (e.g., A100-40GB x2)
srun -p A100-40GB --job-name=train_0.01_1 --ntasks=2 \
    --gpus-per-task=1 --cpus-per-task=3 \
    --mem=60GB --kill-on-bad-exit=1 python -u train.py \
    --cfg_file ./cfgs/kitti_models/pv_rcnn_ssl_60.yaml \
    --batch_size 8 --launcher slurm --split train_0.01_1 \
    --extra_tag train_0.01_1 --ckpt_save_interval 2 \
    --pretrained_model "../output/cfgs/kitti_models/pv_rcnn/split_0.01_1/ckpt/checkpoint_epoch_80.pth" \
    --repeat 5 --dbinfos kitti_dbinfos_train_0.01_1_37.pkl

Acknowledgements

Our code is based on 3DIoUMatch-PVRCNN and OpenPCDet v0.5. Thanks OpenPCDet Development Team for their awesome codebase.

This work has been funded by the German Ministry for Education and Research (BMB+F) in the project MOMENTUM.

Citation

If you find this work useful, please consider citing:

@InProceedings{Nozarian_2023_CVPR,
    author    = {Nozarian, Farzad and Agarwal, Shashank and Rezaeianaran, Farzaneh and Shahzad, Danish and Poibrenski, Atanas and M\"uller, Christian and Slusallek, Philipp},
    title     = {Reliable Student: Addressing Noise in Semi-Supervised 3D Object Detection},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2023},
    pages     = {4980-4989}
}