The following metrics are consistently used in our benchmark:
-
Mean Corruption Error (mCE):
- The Corruption Error (CE) for model
$A$ under corruption type$i$ across 3 severity levels is:$\text{CE}_i^{\text{Model}A} = \frac{\sum((1 - \text{mIoU})^{\text{Model}A})}{\sum((1 - \text{mIoU})^{\text{Baseline}})}$ . - The average CE for model
$A$ on all$N$ corruption types, i.e., mCE, is calculated as:$\text{mCE} = \frac{1}{N}\sum\text{CE}_i$ .
- The Corruption Error (CE) for model
-
Mean Resilience Rate (mRR):
- The Resilience Rate (RR) for model
$A$ under corruption type$i$ across 3 severity levels is:$\text{RR}_i^{\text{Model}A} = \frac{\sum(\text{mIoU}^{\text{Model}A})}{3\times (\text{clean-mIoU}^{\text{Model}A})} .$ - The average RR for model
$A$ on all$N$ corruption types, i.e., mRR, is calculated as:$\text{mRR} = \frac{1}{N}\sum\text{RR}_i$ .
- The Resilience Rate (RR) for model
Corruption | Light | Moderate | Heavy | Average | ||
---|---|---|---|---|---|---|
Fog | 20.21 | 18.62 | 17.72 | 18.85 | 183.89 | 59.63 |
Wet Ground | 29.07 | 26.51 | 26.33 | 27.30 | 158.01 | 86.37 |
Snow | 21.79 | 22.74 | 23.57 | 22.70 | 165.45 | 71.81 |
Motion Blur | 19.52 | 17.86 | 16.42 | 17.93 | 122.35 | 56.72 |
Beam Missing | 28.67 | 24.92 | 21.44 | 25.01 | 171.68 | 79.12 |
Crosstalk | 23.08 | 21.58 | 20.30 | 21.65 | 188.07 | 68.49 |
Incomplete Echo | 28.22 | 27.85 | 26.90 | 27.66 | 158.74 | 87.50 |
Cross-Sensor | 10.45 | 8.07 | 5.04 | 7.85 | 170.81 | 24.83 |
-
Summary:
$\text{mIoU}_{\text{clean}} =$ 31.61%,$\text{mCE} =$ 164.87%,$\text{mRR} =$ 66.81%.
@inproceedings{wu2017squeezeseg,
title = {Squeezeseg: Convolutional Neural Nets with Recurrent CRF for Real-Time Road-Object Segmentation from 3D LiDAR Point Cloud},
author = {Wu, Bichen and Wan, Alvin and Yue, Xiangyu and Keutzer, Kurt},
booktitle = {IEEE International Conference on Robotics and Automation},
year = {2018},
}
@inproceedings{milioto2019rangenet,
title = {RangeNet++: Fast and Accurate LiDAR Semantic Segmentation},
author = {A. Milioto and I. Vizzo and J. Behley and C. Stachniss},
booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
year = {2019},
}