DEIO: Deep Event Inertial Odometry

Paper | Website | Youtube | Bilibili

Abstract

Event cameras are bio-inspired, motion-activated sensors that demonstrate impressive potential in handling challenging situations, such as motion blur and high-dynamic range. Despite their promise, existing event-based simultaneous localization and mapping (SLAM) approaches exhibit limited performance in real-world applications. On the other hand, state-of-the-art SLAM approaches that incorporate deep neural networks for better robustness and applicability. However, these is a lack of research in fusing learning-based event SLAM methods with IMU, which could be indispensable to push the event-based SLAM to large-scale, low-texture or complex scenarios. In this paper, we propose DEIO, the first monocular deep event-inertial odometry framework that combines learning-based method with traditional nonlinear graph-based optimization. Specifically, we tightly integrate a trainable event-based differentiable bundle adjustment (e-DBA) with the IMU pre-integration in a factor graph which employs keyframe-based sliding window optimization. Numerical Experiments in nine public challenge datasets show that our method can achieve superior performance compared with the image-based and event-based benchmarks.

Update log

• README Upload (2024/10/28)
• Paper Upload (2024/11/06)
• Estimated Trajectories Upload (2024/11/07)
• Evaluated Data Upload
• Training Data Upload
• Code Upload (will be released once the paper is accepted)
• Trainable Event Representation

Setup and Installation

Training and Supervision

• The synthetic event-based TartanAir data is generated using the ESIM simulator. Following Note1 and Note2 for quickly use.

Evaluating DEIO

1. DAVIS240C ¹

Download sample sequence from boxes_6dof, poster_translation (ASL format)

Run the DEIO as the following steps:

conda activate deio

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=${YOUR_WORKSPACE} python script/eval_eio/deio_davis240c.py \
--datapath=${YOUR_DATAFOLDER}  \
--weights=eDBA.pth \
--visual_only=0 \
--evaluate_flag \
--enable_event \
--SCORER_EVAL_USE_GRID

Estimated trajectories against the GT in DAVIS240C

2. Mono-HKU Dataset ²

Download sample sequence from vicon_dark1, vicon_hdr4 (ASL format)

Run the DEIO as the following steps:

conda activate deio

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=${YOUR_WORKSPACE} python script/eval_eio/deio_mono_hku.py \
--datapath=${YOUR_DATAFOLDER} \
--weights=eDBA.pth \
--visual_only=0 \
--evaluate_flag \
--enable_event \
--SCORER_EVAL_USE_GRID \
--side=davis346

Estimated trajectories against the GT in Mono-HKU Dataset

3. Stereo-HKU Dataset ³

Download sample sequence from aggressive_translation, hdr_agg (ASL format)

Run the DEIO as the following steps:

conda activate deio

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=${YOUR_WORKSPACE} python script/eval_eio/deio_stereo_hku.py \
--datapath=${YOUR_DATAFOLDER}  \
--weights=eDBA.pth \
--visual_only=0 \
--evaluate_flag \
--enable_event \
--SCORER_EVAL_USE_GRID

Estimated trajectories against the GT in Stereo-HKU Dataset

4. VECtor ⁴

Download sample sequence from corridors_walk1, units_scooter1 (ASL format)

Run the DEIO as the following steps:

conda activate deio

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=${YOUR_WORKSPACE} python script/eval_eio/deio_vector.py \
--datapath=${YOUR_DATAFOLDER}  \
--weights=eDBA.pth \
--visual_only=0 \
--evaluate_flag \
--enable_event \
--SCORER_EVAL_USE_GRID \

Estimated trajectories against the GT in VECtor

5. TUM-VIE ⁵

Download sample sequence from mocap-6dof, mocap-desk2 (ASL format)

Run the DEIO as the following steps:

conda activate deio

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=${YOUR_WORKSPACE} python script/eval_eio/deio_tumvie.py \
--datapath=${YOUR_DATAFOLDER}  \
--weights=eDBA.pth \
--visual_only=0 \
--evaluate_flag \
--enable_event \
--SCORER_EVAL_USE_GRID \

Estimated trajectories against the GT in TUM-VIE Dataset

6. UZH-FPV ⁶

Download sample sequence from indoor_forward_6, indoor_forward_7 (ASL format)

Run the DEIO as the following steps:

conda activate deio

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=${YOUR_WORKSPACE} python script/eval_eio/deio_uzh_fpv.py \
--datapath=${YOUR_DATAFOLDER}  \
--weights=eDBA.pth \
--visual_only=0 \
--evaluate_flag \
--enable_event \
--SCORER_EVAL_USE_GRID \

Estimated trajectories against the GT in UZH-FPV

7. MVSEC ⁷

Download sample sequence from indoor_flying_1, indoor_flying_3 (ASL format)

Run the DEIO as the following steps:

conda activate deio

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=${YOUR_WORKSPACE} python script/eval_eio/deio_mvsec.py \
--datapath=${YOUR_DATAFOLDER}  \
--weights=eDBA.pth \
--visual_only=0 \
--evaluate_flag \
--enable_event \
--SCORER_EVAL_USE_GRID \

Estimated trajectories against the GT in MVSEC Dataset

8. DSEC ⁸

Download sample sequence from dsec_zurich_city_04_a, dsec_zurich_city_04_e (ASL format)

Run the DEIO as the following steps:

conda activate deio

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=${YOUR_WORKSPACE} python script/eval_eio/deio_dsec.py \
--datapath=${YOUR_DATAFOLDER}  \
--weights=eDBA.pth \
--visual_only=0 \
--evaluate_flag \
--enable_event \
--SCORER_EVAL_USE_GRID \

Estimated trajectories against the GT in DSEC

Run on Your Own Dataset

• Taking ECMD ⁹ as an example. First download the rosbag file, and then run the following command:

conda activate deio

CUDA_VISIBLE_DEVICES=2 PYTHONPATH=${YOUR_WORKSPACE} python script/pp_data/pp_ecmd.py --indir=${YOUR_DATAFOLDER}

• Duplicate a script from deio_davis240c.py or deio_ecmd.py
• In the script, specify the data loading procedure of IMU data and Event loader.
• Specify the timestamp file and unit for both event streams and IMU.
• Specify the event camera intrinsics and camera-IMU extrinsics in the script.
• Try it!

Estimated trajectories of our DEIO against the GNSS-INS-RTK in ECMD

Using Our Results as Comparison

For the convenience of the comparison, we release the estimated trajectories of DEIO in tum format in the dir of estimated_trajectories. What's more, we also give the sample code for the quantitative and qualitative evaluation using evo package

Acknowledgement

• This work is based on DPVO, DEVO, DROID-SLAM, DBA-Fusion, and GTSAM
• More details about the trainable event representation is available in
• If you find this work is helpful in your research, a simple star or citation of our works should be the best affirmation for us. 😊

@article{GWPHKU:DEIO,
  title={DEIO: Deep Event Inertial Odometry},
  author={Guan, Weipeng and Lin, Fuling and Chen, Peiyu and Lu, Peng},
  journal={arXiv preprint arXiv:2411.03928},
  year={2024}
}