Not Every Side Is Equal: Localization Uncertainty Estimation for Semi-Supervised 3D Object Detection

This repository contains an implementation of Nesie, a Semi-Supervised 3D Object Detection method introduced in our paper:

[Project Webpage] [Paper]

News

• 18. July 2023: Nesie is accepted at ICCV 2023.
• October 2023: Nesie PDF released.
• December 2023: Code (Nesie) for ScanNet dataset released.
• May 2024: Code (SAQE) for ScanNet dataset released. 🔥

Todo

• [✔] Code (Nesie) for ScanNet dataset.
• [✔] Code (SAQE) for ScanNet dataset.
• Code (Nesie) for SunRGB-D dataset.
• Code (SAQE) for SunRGB-D dataset.

Installation

Please following the env_setup.sh

Getting Started

We follow the mmdetection3d data preparation protocol described in scannet, sunrgbd.

Pre-training

To start pre-training, run with Nesie configs or SAQE configs:

CUDA_VISIBLE_DEVICES=$gpu_id OMP_NUM_THREADS=24 \
python tools/train.py \
configs/Nesie(SAQE)/nesie(saqe)-votenet-scannet-pretrain-$Ratio.py \
--gpu-ids 0 

Training

To start training, run with Nesie configs or SAQE configs:

CUDA_VISIBLE_DEVICES=$gpu_id OMP_NUM_THREADS=24 \
python tools/train.py \
configs/Nesie(SAQE)/nesie(saqe)-votenet-scannet-train-$Ratio.py \
--gpu-ids 0 \
--load-from work_dirs/nesie(saqe)-votenet-scannet-pretrain-$Ratio/epoch_36.pth

Testing

Test model using with Nesie configs or SAQE configs:

CUDA_VISIBLE_DEVICES=$gpu_id OMP_NUM_THREADS=24 \
python tools/test.py \
configs/Nesie(SAQE)/nesie(saqe)-votenet-scannet-test.py \
work_dirs/nesie(saqe)-votenet-scannet-train-$Ratio/epoch_36.pth --eval mAP --seed 9

Results

Comparison with state-of-the-art methods on ScanNet dataset. Ours† represents the Nesie version and Ours represents the SAQE version.

Comparison with state-of-the-art methods on SunRGBD dataset. Ours† represents the Nesie version and Ours represents the SAQE version.

Visulization of Detection Results

Citation

If you find this work useful for your research, please cite our paper:

@InProceedings{Wang_2023_ICCV,
    author    = {ChuXin Wang and Wenfei Yang and Tianzhu Zhang},
    title     = {Not Every Side Is Equal: Localization Uncertainty Estimation for Semi-Supervised 3D Object Detection},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
}