Skip to content

Latest commit

 

History

History
154 lines (110 loc) · 9.16 KB

Sparse4D-r101.md

File metadata and controls

154 lines (110 loc) · 9.16 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:

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

Sparse4D R101

Corruption NDS mAP mATE mASE mAOE mAVE mAAE
Clean 0.5438 0.4409 0.6282 0.2721 0.3853 0.2922 0.1888
Cam Crash 0.2873 0.1319 0.7852 0.2917 0.4989 0.9611 0.2510
Frame Lost 0.2611 0.1050 0.8175 0.3166 0.5404 1.0253 0.2726
Color Quant 0.3310 0.2345 0.8348 0.2956 0.5452 0.9712 0.2496
Motion Blur 0.2514 0.1438 0.8719 0.3553 0.6780 1.0817 0.3347
Brightness 0.3984 0.3296 0.7543 0.2835 0.4844 0.9232 0.2187
Low Light 0.2510 0.1386 0.8501 0.3543 0.6464 1.1621 0.3356
Fog 0.3884 0.3097 0.7552 0.2840 0.4933 0.9087 0.2229
Snow 0.2259 0.1275 0.8860 0.3875 0.7116 1.1418 0.3936

Experiment Log

Time: Day Month xx xx:xx:xx 2023

Camera Crash

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.3369 0.2035 0.7533 0.2857 0.4744 0.9023 0.2325
Moderate 0.2623 0.0979 0.8150 0.2933 0.5016 1.0038 0.2571
Hard 0.2628 0.0944 0.7874 0.2962 0.5206 0.9772 0.2633
Average 0.2873 0.1319 0.7852 0.2917 0.4989 0.9611 0.2510

Frame Lost

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.3494 0.2259 0.7498 0.2847 0.4649 0.9060 0.2306
Moderate 0.2479 0.0746 0.8204 0.3008 0.5351 0.9929 0.2449
Hard 0.1861 0.0143 0.8824 0.3643 0.6212 1.1770 0.3423
Average 0.2611 0.1050 0.8175 0.3166 0.5404 1.0253 0.2726

Color Quant

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.4109 0.3427 0.7488 0.2801 0.4620 0.8837 0.2295
Moderate 0.3385 0.2462 0.8222 0.2899 0.5613 0.9295 0.2433
Hard 0.2435 0.1147 0.9333 0.3167 0.6124 1.1004 0.2761
Average 0.3310 0.2345 0.8348 0.2956 0.5452 0.9712 0.2496

Motion Blur

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.3692 0.2830 0.7731 0.2898 0.5352 0.8952 0.2292
Moderate 0.2169 0.0927 0.9302 0.3286 0.7464 1.0806 0.2895
Hard 0.1681 0.0557 0.9124 0.4475 0.7525 1.2693 0.4853
Average 0.2514 0.1438 0.8719 0.3553 0.6780 1.0817 0.3347

Brightness

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.4273 0.3626 0.7157 0.2801 0.4501 0.8722 0.2219
Moderate 0.3991 0.3321 0.7538 0.2838 0.4918 0.9238 0.2157
Hard 0.3687 0.2942 0.7933 0.2866 0.5114 0.9737 0.2186
Average 0.3984 0.3296 0.7543 0.2835 0.4844 0.9232 0.2187

Low Light

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.3115 0.2096 0.8259 0.3005 0.5749 0.9894 0.2419
Moderate 0.2613 0.1398 0.8561 0.3154 0.6396 1.1260 0.2751
Hard 0.1803 0.0664 0.8682 0.4471 0.7246 1.3708 0.4897
Average 0.2510 0.1386 0.8501 0.3543 0.6464 1.1621 0.3356

Fog

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.4021 0.3294 0.7434 0.2836 0.4893 0.8918 0.2179
Moderate 0.3926 0.3128 0.7516 0.2828 0.4819 0.8990 0.2228
Hard 0.3706 0.2869 0.7706 0.2857 0.5087 0.9353 0.2280
Average 0.3884 0.3097 0.7552 0.2840 0.4933 0.9087 0.2229

Snow

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.2259 0.1275 0.8860 0.3875 0.7116 1.1418 0.3936
Moderate 0.1757 0.0736 0.9395 0.4159 0.8436 1.2643 0.4121
Hard 0.1682 0.0654 0.9548 0.4160 0.8515 1.2578 0.4229
Average 0.1899 0.0888 0.9268 0.4065 0.8022 1.2213 0.4095

References

@article{lin2022sparse4d,
  title={Sparse4D: Multi-view 3D Object Detection with Sparse Spatial-Temporal Fusion},
  author={Lin, Xuewu and Lin, Tianwei and Pei, Zixiang and Huang, Lichao and Su, Zhizhong},
  journal={arXiv preprint arXiv:2211.10581},
  year={2022}
}