Ev-3DOD (CVPR 2025 Highlight)
August 2, 2025 ยท View on GitHub
We would like to begin by expressing our sincere gratitude to the reviewers and the area chairs for their valuable feedback and for recognizing the contributions of our paper.
:star2: Update (02/08/2025) :star2: We have released Ev-Waymo dataset on Hugging Face !
:star2: Update (20/07/2025) :star2: We have released DSEC-3DOD dataset on Hugging Face !
:star2: Update (09/06/2025) :star2: We have released train and evaluation codes !
:star2: Update (31/12/2024) :star2: We have released Ev-Waymo and DSEC-3DOD datasets !
This repo is the official implementation of: Ev-3DOD: Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras, CVPR 2025.
Paper
:video_camera: Demo Video
A demo video Youtube can be found by clicking the image below:
Framework
Ev-3DOD consists of two training stages.
- In Stage 1, the model follows a conventional box proposal approach without using event data.
- In Stage 2, event data is fused with other sensor modalities to enable detection during blind time intervals.
Downloading Ev-Waymo and DSEC-3DOD Datasets
In this study, we used the following two datasets. In particular, we manually annotated the DSEC-3DOD dataset. If you use the following datasets, please cite this work and also cite the original papers corresponding to the datasets.
-
Ev-Waymo dataset can be downloaded from the link hugging face.
-
DSEC-3DOD dataset can be downloaded from the google drive or hugging face.
Running
Please read the corresponding README for details
Stage1 (only for 10FPS (not using Events))
Stage2 (100FPS inference (using Events))
Main results
Results on Ev-Waymo Dataset
Results on DSEC-3DOD Dataset
Our main paper reports results using the metrics provided by the Waymo Open Dataset. For broader comparability and to facilitate future research, we also re-evaluated the same models using the KITTI metrics. The results can be found at the link below.
Acknowledgement
We sincerely appreciate the following open-source projects for providing valuable and high-quality codes:
- DSEC
- LoGoNet
- mmdetection3d
- CenterPoint
- BEVFusion(ADLab-AutoDrive)
- BEVFusion(mit-han-lab)
- mmdetection
- PDV
Reference
If you find our paper useful, please kindly cite us via:
@inproceedings{cho2025ev,
title={Ev-3dod: Pushing the temporal boundaries of 3d object detection with event cameras},
author={Cho, Hoonhee and Kang, Jae-young and Kim, Youngho and Yoon, Kuk-Jin},
booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
pages={27197--27210},
year={2025}
}
Contact
- If you have any questions regarding datasets and codes, please leave an issue or contact mickeykang@kaist.ac.kr, gnsgnsgml@kaist.ac.kr.