Skip to content

Latest commit

 

History

History
156 lines (111 loc) · 10 KB

BEVDet-r101-FCOS3D-Pretrain.md

File metadata and controls

156 lines (111 loc) · 10 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:

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

BEVDet-r101-fcos3d-pretrain

Corruption NDS mAP mATE mASE mAOE mAVE mAAE
Clean 0.3780 0.2846 0.7274 0.2796 0.5517 0.8581 0.2264
Cam Crash 0.2442 0.0928 0.8020 0.3384 0.5815 1.0285 0.3453
Frame Lost 0.1962 0.0720 0.8320 0.4427 0.6830 1.0063 0.4684
Color Quant 0.3041 0.2064 0.7815 0.3247 0.6251 0.9955 0.3212
Motion Blur 0.2590 0.1512 0.7826 0.3675 0.6412 1.1481 0.3973
Brightness 0.2599 0.1714 0.7910 0.3963 0.6828 1.1539 0.4242
Low Light 0.1393 0.0613 0.8761 0.5631 0.8235 1.1739 0.6510
Fog 0.2073 0.0984 0.8521 0.4107 0.6897 1.2659 0.4668
Snow 0.0939 0.0301 0.9494 0.6685 0.8397 1.2412 0.7535

Experiment Log

Time: Mon Mar 6 22:08:01 2023

Camera Crash

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.3036 0.1480 0.7628 0.2800 0.5572 0.8635 0.2410
Moderate 0.2253 0.0629 0.8272 0.3323 0.5746 1.0089 0.3277
Hard 0.2039 0.0675 0.8161 0.4030 0.6128 1.2132 0.4672
Average 0.2442 0.0928 0.8020 0.3384 0.5815 1.0285 0.3453

Frame Lost

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.3058 0.1651 0.7656 0.2813 0.5715 0.9164 0.2329
Moderate 0.2142 0.0449 0.8123 0.3347 0.6042 1.0881 0.3317
Hard 0.0685 0.0059 0.9182 0.7122 0.8732 1.0145 0.8407
Average 0.1962 0.0720 0.8320 0.4427 0.6830 1.0063 0.4684

Color Quant

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.3726 0.2790 0.7226 0.2784 0.5556 0.8760 0.2365
Moderate 0.3299 0.2296 0.7658 0.2799 0.6044 0.9406 0.2579
Hard 0.2096 0.1105 0.8561 0.4157 0.7154 1.1700 0.4691
Average 0.3041 0.2064 0.7815 0.3247 0.6251 0.9955 0.3212

Motion Blur

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.3504 0.2502 0.7417 0.2795 0.5642 0.9306 0.2308
Moderate 0.2498 0.1242 0.7825 0.3365 0.6505 1.1741 0.3532
Hard 0.1768 0.0790 0.8235 0.4864 0.7090 1.3395 0.6078
Average 0.2590 0.1512 0.7826 0.3675 0.6412 1.1481 0.3973

Brightness

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.3539 0.2532 0.7359 0.2794 0.5774 0.8897 0.2445
Moderate 0.2445 0.1607 0.7900 0.4148 0.7054 1.2555 0.4482
Hard 0.1813 0.1001 0.8470 0.4947 0.7657 1.3165 0.5799
Average 0.2599 0.1714 0.7910 0.3963 0.6828 1.1539 0.4242

Low Light

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.2143 0.1077 0.8034 0.4097 0.7081 1.2486 0.4745
Moderate 0.1588 0.0602 0.8594 0.4910 0.7623 1.1715 0.5996
Hard 0.0447 0.0160 0.9656 0.7886 1.0000 1.1016 0.8789
Average 0.1393 0.0613 0.8761 0.5631 0.8235 1.1739 0.6510

Fog

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.2546 0.1356 0.8267 0.3305 0.6306 1.1996 0.3441
Moderate 0.2072 0.0972 0.8393 0.4128 0.6949 1.3338 0.4669
Hard 0.1600 0.0625 0.8902 0.4888 0.7437 1.2643 0.5895
Average 0.2073 0.0984 0.8521 0.4107 0.6897 1.2659 0.4668

Snow

Severity NDS mAP mATE mASE mAOE mAVE mAAE
Easy 0.1691 0.0543 0.8868 0.4269 0.7774 1.4151 0.4896
Moderate 0.0804 0.0275 0.9744 0.7199 0.7926 1.1624 0.8465
Hard 0.0323 0.0085 0.9870 0.8587 0.9490 1.1461 0.9245
Average 0.0939 0.0301 0.9494 0.6685 0.8397 1.2412 0.7535

References

@article{huang2021bevdet,
  title={Bevdet: High-performance multi-camera 3d object detection in bird-eye-view},
  author={Huang, Junjie and Huang, Guan and Zhu, Zheng and Du, Dalong},
  journal={arXiv preprint arXiv:2112.11790},
  year={2021}
}
}