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 | 62.15 | 59.05 | 48.39 | 56.53 | 98.50 | 89.42 |
Wet Ground | 59.43 | 51.83 | 49.79 | 53.68 | 100.67 | 84.91 |
Snow | 54.32 | 52.22 | 50.52 | 52.35 | 101.99 | 82.81 |
Motion Blur | 46.67 | 32.10 | 24.40 | 34.39 | 97.81 | 54.40 |
Beam Missing | 60.66 | 57.36 | 52.26 | 56.76 | 98.99 | 89.78 |
Crosstalk | 60.56 | 59.22 | 57.23 | 59.00 | 98.42 | 93.32 |
Incomplete Echo | 58.41 | 54.90 | 51.61 | 54.97 | 98.82 | 86.95 |
Cross-Sensor | 57.72 | 52.72 | 30.76 | 47.07 | 98.11 | 74.45 |
-
Summary:
$\text{mIoU}_{\text{clean}} =$ 63.22%,$\text{mCE} =$ 99.16%,$\text{mRR} =$ 82.01%.
@inproceedings{tang2020searching,
title = {Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution},
author = {Tang, Haotian and Liu, Zhijian and Zhao, Shengyu and Lin, Yujun and Lin, Ji and Wang, Hanrui and Han, Song},
booktitle = {European Conference on Computer Vision}
year = {2020}
}