Unleashing the Temporal Potential of Stereo Event Cameras for Continuous-Time 3D Object Detection

Jae-Young Kang^*  ·  Hoonhee Cho^*  ·  Kuk-Jin Yoon

KAIST

ICCV 2025

Paper | Code

Introduction

• We present the first continuous-time 3D object detection framework that relies solely on stereo event cameras, eliminating the need for LiDAR or RGB sensors.
• A dual-filter mechanism extracts both semantic and geometric information from event data, while object-centric regression improves bounding box alignment.
• Experiments show superior performance in fast and dynamic environments, proving the potential of event cameras for robust continuous-time 3D perception.

Overview

• Installation
• Getting Started
• Pretrained Models
• Citation

Installation

• Tested on Cuda 11.1

Clone this repository.

git clone https://github.com/mickeykang16/Ev-Stereo3D.git

Set up new environment

conda create -n evstereo3d python=3.7 -y
conda activate evstereo3d

Install the dependent libraries as follows:

pip install torch==1.10.0+cu111 torchvision==0.11.0+cu111 torchaudio==0.10.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install -r requirements.txt 

• Install the SparseConv library, we use the implementation from [spconv].

git clone https://github.com/traveller59/spconv
git reset --hard f22dd9
git submodule update --recursive
python setup.py bdist_wheel
pip install ./dist/spconv-1.2.1-cp37-cp37m-linux_x86_64.whl

• Install modified mmdetection

cd mmdetection_kitti
python setup.py develop

• Install modules from [LoGoNet]

git clone https://github.com/PJLab-ADG/LoGoNet.git
cd LoGoNet/utils && python setup.py develop
cd LoGoNet/detection  && python setup.py develop

• Install this library by running the following command:

python setup.py develop

Getting Started

The dataset configs are located within configs/stereo/dataset_configs, and the model configs are located within configs/stereo for different datasets.

Dataset Preparation

Please refer to DSEC dataset preparation

Training & Testing

Test and evaluate the pretrained models (10FPS)

• To test with multiple GPUs:

./scripts/dist_test_ckpt.sh ${GPU_IDS} ./configs/stereo/dsec_models/evstereo.yaml ./ckpt/checkpoint_epoch_22.pth

Train a model

• Train with multiple GPUs
• The code only supports batch size of 1 per GPU

./scripts/dist_train.sh ${GPU_IDS} ${EXP_NAME} ./configs/stereo/dsec_models/evstereo.yaml --find_unused_parameters

Citation

@InProceedings{Kang_2025_ICCV,
    author    = {Kang, Jae-Young and Cho, Hoonhee and Yoon, Kuk-Jin},
    title     = {Unleashing the Temporal Potential of Stereo Event Cameras for Continuous-Time 3D Object Detection},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2025},
    pages     = {6869-6881}
}

Acknowledgements

Part of codes are migrated from LIGA Stereo, OpenPCDet and DSGN.