Skip to content

Latest commit

 

History

History
155 lines (110 loc) · 10.1 KB

SRCN3D-r101.md

File metadata and controls

155 lines (110 loc) · 10.1 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:

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

SRCN3D-R101

Corruption NDS mAP mATE mASE mAOE mAVE mAAE
Clean 0.4286 0.3373 0.7783 0.2873 0.3665 0.7806 0.1878
Cam Crash 0.2947 0.1172 0.8369 0.3017 0.4403 0.8506 0.2097
Frame Lost 0.2681 0.0924 0.8637 0.3303 0.4798 0.8725 0.2349
Color Quant 0.3318 0.2199 0.8696 0.3041 0.4747 0.8877 0.2458
Motion Blur 0.2609 0.1361 0.9026 0.3524 0.5788 0.9964 0.2927
Brightness 0.4074 0.3133 0.7936 0.2911 0.3974 0.8227 0.1877
Low Light 0.2590 0.1406 0.8586 0.3642 0.5773 1.1257 0.3353
Fog 0.3940 0.2932 0.7993 0.2919 0.3978 0.8428 0.1944
Snow 0.1920 0.0734 0.9372 0.3996 0.7302 1.2366 0.3803

Experiment Log

Time: Mon Feb 20 13:17:16 2023

Camera Crash

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.3399 0.1823 0.8094 0.2918 0.3864 0.8221 0.2024
Moderate 0.2633 0.0810 0.8743 0.3031 0.4362 0.9321 0.2260
Hard 0.2808 0.0884 0.8270 0.3101 0.4984 0.7976 0.2007
Average 0.2947 0.1172 0.8369 0.3017 0.4403 0.8506 0.2097

Frame Lost

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.3513 0.2016 0.8095 0.2907 0.3917 0.8144 0.1888
Moderate 0.2604 0.0643 0.8618 0.3110 0.4746 0.8643 0.2055
Hard 0.1925 0.0113 0.9199 0.3892 0.5732 0.9389 0.3103
Average 0.2681 0.0924 0.8637 0.3303 0.4798 0.8725 0.2349

Color Quant

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.4088 0.3171 0.7977 0.2880 0.3850 0.8204 0.2064
Moderate 0.3458 0.2365 0.8506 0.2983 0.4660 0.8748 0.2347
Hard 0.2407 0.1060 0.9605 0.3260 0.5731 0.9678 0.2962
Average 0.3318 0.2199 0.8696 0.3041 0.4747 0.8877 0.2458

Motion Blur

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.3806 0.2754 0.8041 0.2932 0.4251 0.8358 0.2128
Moderate 0.2249 0.0861 0.9392 0.3454 0.6254 1.0136 0.2716
Hard 0.1772 0.0470 0.9646 0.4186 0.6858 1.1397 0.3938
Average 0.2609 0.1361 0.9026 0.3524 0.5788 0.9964 0.2927

Brightness

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.4250 0.3331 0.7815 0.2879 0.3760 0.7843 0.1864
Moderate 0.4079 0.3150 0.7942 0.2906 0.3965 0.8273 0.1875
Hard 0.3893 0.2916 0.8051 0.2947 0.4196 0.8565 0.1892
Average 0.4074 0.3133 0.7936 0.2911 0.3974 0.8227 0.1877

Low Light

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.3252 0.2111 0.8360 0.3017 0.4853 0.9329 0.2478
Moderate 0.2707 0.1455 0.8588 0.3236 0.5625 1.0326 0.2762
Hard 0.1812 0.0653 0.8809 0.4672 0.6840 1.4115 0.4820
Average 0.2590 0.1406 0.8586 0.3642 0.5773 1.1257 0.3353

Fog

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.4081 0.3101 0.7893 0.2894 0.3859 0.8155 0.1892
Moderate 0.3976 0.2960 0.7980 0.2909 0.3906 0.8330 0.1919
Hard 0.3763 0.2736 0.8106 0.2953 0.4169 0.8800 0.2021
Average 0.3940 0.2932 0.7993 0.2919 0.3978 0.8428 0.1944

Snow

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.2503 0.1200 0.9046 0.3344 0.5726 1.1201 0.2858
Moderate 0.1706 0.0552 0.9387 0.4265 0.7846 1.2776 0.4201
Hard 0.1550 0.0450 0.9682 0.4380 0.8334 1.3122 0.4349
Average 0.1920 0.0734 0.9372 0.3996 0.7302 1.2366 0.3803

References

@article{shi2022srcn3d,
  title={Srcn3d: Sparse r-cnn 3d surround-view camera object detection and tracking for autonomous driving},
  author={Shi, Yining and Shen, Jingyan and Sun, Yifan and Wang, Yunlong and Li, Jiaxin and Sun, Shiqi and Jiang, Kun and Yang, Diange},
  journal={arXiv preprint arXiv:2206.14451},
  year={2022}
}