Skip to content

Latest commit

 

History

History
131 lines (94 loc) · 7.91 KB

BEVFusion-Fusion.md

File metadata and controls

131 lines (94 loc) · 7.91 KB

RoboBEV Benchmark

The official nuScenes metrics are considered in our benchmark:

Average Precision (AP)

The average precision (AP) defines a match by thresholding the 2D center distance d on the ground plane instead of the intersection over union (IoU). This is done in order to decouple detection from object size and orientation but also because objects with small footprints, like pedestrians and bikes, if detected with a small translation error, give $0$ IoU. We then calculate AP as the normalized area under the precision-recall curve for recall and precision over 10%. Operating points where recall or precision is less than $10$% are removed in order to minimize the impact of noise commonly seen in low precision and recall regions. If no operating point in this region is achieved, the AP for that class is set to zero. We then average over-matching thresholds of $\mathbb{D}={0.5, 1, 2, 4}$ meters and the set of classes $\mathbb{C}$ :

$$ \text{mAP}= \frac{1}{|\mathbb{C}||\mathbb{D}|}\sum_{c\in\mathbb{C}}\sum_{d\in\mathbb{D}}\text{AP}_{c,d} . $$

True Positive (TP)

All TP metrics are calculated using $d=2$ m center distance during matching, and they are all designed to be positive scalars. Matching and scoring happen independently per class and each metric is the average of the cumulative mean at each achieved recall level above $10$%. If a $10$% recall is not achieved for a particular class, all TP errors for that class are set to $1$.

  • Average Translation Error (ATE) is the Euclidean center distance in 2D (units in meters).
  • Average Scale Error (ASE) is the 3D intersection-over-union (IoU) after aligning orientation and translation ($1$ − IoU).
  • Average Orientation Error (AOE) is the smallest yaw angle difference between prediction and ground truth (radians). All angles are measured on a full $360$-degree period except for barriers where they are measured on a $180$-degree period.
  • Average Velocity Error (AVE) is the absolute velocity error as the L2 norm of the velocity differences in 2D (m/s).
  • Average Attribute Error (AAE) is defined as $1$ minus attribute classification accuracy ($1$ − acc).

nuScenes Detection Score (NDS)

mAP with a threshold on IoU is perhaps the most popular metric for object detection. However, this metric can not capture all aspects of the nuScenes detection tasks, like velocity and attribute estimation. Further, it couples location, size, and orientation estimates. nuScenes proposed instead consolidating the different error types into a scalar score:

$$ \text{NDS} = \frac{1}{10} [5\text{mAP}+\sum_{\text{mTP}\in\mathbb{TP}} (1-\min(1, \text{mTP}))] . $$

BEVFusion Camera + LiDAR

Corruption NDS mAP mATE mASE mAOE mAVE mAAE
Clean 0.7138 0.6852 0.2874 0.2539 0.3044 0.2554 0.1874
Cam Crash 0.6963 0.6525 0.2916 0.2565 0.3069 0.2603 0.1847
Frame Lost 0.6931 0.6478 0.2927 0.2573 0.3112 0.2628 0.1840
Color Quant 0.7044 0.6665 0.2886 0.2553 0.2980 0.2569 0.1891
Motion Blur 0.6977 0.6557 0.2885 0.2560 0.3116 0.2598 0.1853
Brightness 0.7018 0.6622 0.2926 0.2566 0.2996 0.2593 0.1852
Low Light 0.6787 0.6210 0.2939 0.2600 0.2998 0.2739 0.1898
Fog - - - - - - -
Snow - - - - - - -

Experiment Log

Time: Thu Apr 6 11:46:46 2023

Camera Crash

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.7044 0.6667 0.2883 0.2560 0.3040 0.2570 0.1841
Moderate 0.6946 0.6511 0.2945 0.2567 0.3082 0.2644 0.1855
Hard 0.6898 0.6398 0.2920 0.2569 0.3086 0.2594 0.1846
Average 0.6963 0.6525 0.2916 0.2565 0.3069 0.2603 0.1847

Frame Lost

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.7027 0.6643 0.2904 0.2557 0.3049 0.2580 0.1852
Moderate 0.6903 0.6438 0.2946 0.2575 0.3141 0.2659 0.1835
Hard 0.6862 0.6354 0.2932 0.2587 0.3146 0.2646 0.1832
Average 0.6931 0.6478 0.2927 0.2573 0.3112 0.2628 0.1840

Color Quant

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.7122 0.6818 0.2876 0.2544 0.3000 0.2557 0.1889
Moderate 0.7076 0.6714 0.2889 0.2556 0.2925 0.2560 0.1880
Hard 0.6934 0.6461 0.2894 0.2558 0.3016 0.2591 0.1904
Average 0.7044 0.6665 0.2886 0.2553 0.2980 0.2569 0.1891

Motion Blur

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.7091 0.6765 0.2867 0.2545 0.3117 0.2544 0.1840
Moderate 0.6952 0.6500 0.2881 0.2563 0.3083 0.2594 0.1858
Hard 0.6889 0.6406 0.2906 0.2572 0.3148 0.2655 0.1860
Average 0.6977 0.6557 0.2885 0.2560 0.3116 0.2598 0.1853

Brightness

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.7082 0.6760 0.2918 0.2556 0.3040 0.2593 0.1869
Moderate 0.7000 0.6585 0.2932 0.2567 0.2989 0.2596 0.1840
Hard 0.6971 0.6521 0.2927 0.2576 0.2958 0.2590 0.1846
Average 0.7018 0.6622 0.2926 0.2566 0.2996 0.2593 0.1852

Low Light

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.6875 0.6360 0.2923 0.2589 0.2989 0.2647 0.1897
Moderate 0.6805 0.6232 0.2925 0.2594 0.2972 0.2715 0.1897
Hard 0.6682 0.6038 0.2970 0.2617 0.3033 0.2855 0.1900
Average 0.6787 0.6210 0.2939 0.2600 0.2998 0.2739 0.1898

References

@inproceedings{liu2022bevfusion,
  title={BEVFusion: Multi-Task Multi-Sensor Fusion with Unified Bird's-Eye View Representation},
  author={Liu, Zhijian and Tang, Haotian and Amini, Alexander and Yang, Xingyu and Mao, Huizi and Rus, Daniela and Han, Song},
  booktitle={IEEE International Conference on Robotics and Automation (ICRA)},
  year={2023}
}