FAST-LIVO2 icon indicating copy to clipboard operation
FAST-LIVO2 copied to clipboard

New Feature: Support for NTU-VIRAL Dataset

Open xuankuzcr opened this issue 8 months ago • 8 comments

🎉 We’ve added support for the NTU-VIRAL dataset! Here’s what’s included:

  • Launch file and unified parameters for all sequences.
  • A more user-friendly evaluation method handling PRISM-IMU extrinsic parameters.
  • RMSE improved from ~6cm (FAST-LIVO2 paper) to ~2cm, outperforming VIRAL-SLAM, BALM2, etc. (best result in SLAM systems to date).
  • Runtime to data duration ratio is 1:6.

我们新增了对NTU-VIRAL数据集的支持!具体更新如下:

  • 提供了launch文件和适用于所有序列的统一参数
  • 提供了一套更友好的评估方法,考虑了棱镜与IMU的外参。
  • RMSE从约 6cm (FAST-LIVO2论文中的实验结果) 提升至约 2cm (当前开源版本),优于VIRAL-SLAM和BALM2等 (是目前据我所知SLAM系统中的最佳结果)。
  • 耗时与数据时长的比例为1:6。
Scene Sequence RMSE (cm)
eee 01 2.71
eee 02 2.11
eee 03 2.61
nya 01 3.56
nya 02 3.39
nya 03 3.52
sbs 01 2.34
sbs 02 2.83
sbs 03 3.11

xuankuzcr avatar Mar 29 '25 10:03 xuankuzcr

Image

xuankuzcr avatar Mar 29 '25 13:03 xuankuzcr

太厉害了郑博!附上论文里面其他模型的精度:(单位:m)

Image

论文链接: https://arxiv.org/pdf/2202.00379

NTU_VIRAL dataset链接: https://ntu-aris.github.io/ntu_viral_dataset/

1391741823 avatar Mar 30 '25 03:03 1391741823

Hello and thank you for your excellent work. When running the data eee01.bag with the configuration file you provided, I can only achieve an accuracy of 2.4-2.6 dm after multiple attempts, how do I need to adjust it to achieve the best accuracy?

Image

Image

JingYCB avatar Apr 08 '25 01:04 JingYCB

Hello, you probably didn't consider the extrinsic conversion between IMU and PRISM coordinate systems before comparing the accuracy. Please refer to the author's link for details: https://github.com/hku-mars/FAST-LIVO2/blob/main/Log/result/ntu_viral/evaluate_viral.py

Xujiaming1 avatar Apr 08 '25 07:04 Xujiaming1

@JingYCB Please follow the evaluation guidance.

xuankuzcr avatar Apr 08 '25 08:04 xuankuzcr

Hello, you probably didn't consider the extrinsic conversion between IMU and PRISM coordinate systems before comparing the accuracy. Please refer to the author's link for details: https://github.com/hku-mars/FAST-LIVO2/blob/main/Log/result/ntu_viral/evaluate_viral.py Thank you very much, resolved

JingYCB avatar Apr 09 '25 07:04 JingYCB

@JingYCB Please follow the evaluation guidance. Thank you very much, resolved

JingYCB avatar Apr 09 '25 07:04 JingYCB

When running fastlio2, do we also need to consider the extrinsic conversion between IMU and PRISM coordinate systems before comparing the accuracy? Further, under what circumstances do we need to convert to this coordinate system? Thank you very much.

JingYCB avatar Apr 09 '25 13:04 JingYCB