/RoboBEV

RoboBEV: Towards Robust Bird's Eye View Perception under Common Corruption and Domain Shift

Primary LanguagePython

English | 简体中文

Benchmarking and Improving Bird's Eye View Perception Robustness
in Autonomous Driving

Shaoyuan Xie1   Lingdong Kong2,3   Wenwei Zhang2,4   Jiawei Ren4   Liang Pan2   Kai Chen2   Ziwei Liu4
1University of California, Irvine   2Shanghai AI Laboratory   3National University of Singapore   4S-Lab, Nanyang Technological University

About

RoboBEV is the first robustness evaluation benchmark tailored for camera-based bird's eye view (BEV) perception under natural data corruption and domain shift, which are cases that have a high likelihood to occur in real-world deployments.

[Common Corruption] - We investigate eight data corruption types that are likely to appear in driving scenarios, ranging from 1sensor failure, 2motion & data processing, 3lighting conditions, and 4weather conditions.

[Domain Shift] - We benchmark the adaptation performance of BEV models from three aspects, including 1city-to-city, 2day-to-night, and 3dry-to-rain.
FRONT_LEFT FRONT FRONT_RIGHT FRONT_LEFT FRONT FRONT_RIGHT
BACK_LEFT BACK BACK_RIGHT BACK_LEFT BACK BACK_RIGHT

Visit our project page to explore more examples. 🚙

Updates

  • [2024.06] - Check out our updated paper for robust BEV perception: Benchmarking and Improving Bird's Eye View Perception Robustness in Autonomous Driving. ⛽
  • [2024.05] - Check out the technical report of this competition: The RoboDrive Challenge: Drive Anytime Anywhere in Any Condition. 🚙
  • [2024.05] - The slides of the 2024 RoboDrive Workshop are available here ⤴️.
  • [2024.05] - The video recordings are available on YouTube ⤴️ and Bilibili ⤴️.
  • [2024.05] - We are glad to announce the winning teams of the 2024 RoboDrive Challenge:
    • Track 1: Robust BEV Detection
      • 🥇 DeepVision, 🥈 Ponyville Autonauts Ltd, 🥉 CyberBEV
    • Track 2: Robust Map Segmentation
      • 🥇 SafeDrive-SSR, 🥈 CrazyFriday, 🥉 Samsung Research
    • Track 3: Robust Occupancy Prediction
      • 🥇 ViewFormer, 🥈 APEC Blue, 🥉 hm.unilab
    • Track 4: Robust Depth Estimation
      • 🥇 HIT-AIIA, 🥈 BUAA-Trans, 🥉 CUSTZS
    • Track 5: Robust Multi-Modal BEV Detection
      • 🥇 safedrive-promax, 🥈 Ponyville Autonauts Ltd, 🥉 HITSZrobodrive
  • [2024.01] - The toolkit tailored for the 2024 RoboDrive Challenge has been released. 🛠️
  • [2023.12] - We are hosting the RoboDrive Challenge at ICRA 2024. 🚙
  • [2023.06] - The nuScenes-C dataset is now available at OpenDataLab! 🚀
  • [2023.04] - We establish "Robust BEV Perception" leaderboards on Paper-with-Code. Join the challenge today! 🙋
  • [2023.02] - We invite every BEV enthusiast to participate in the robust BEV perception benchmark! For more details, please read this page. 🍻
  • [2023.01] - Launch of RoboBEV! In this initial version, 11 BEV detection algorithms and 1 monocular 3D detection algorithm have been benchmarked under 8 corruption types across 3 severity levels.

Outline

Installation

Kindly refer to INSTALL.md for the installation details.

Data Preparation

Our datasets are hosted by OpenDataLab.


OpenDataLab is a pioneering open data platform for the large AI model era, making datasets accessible. By using OpenDataLab, researchers can obtain free formatted datasets in various fields.

Kindly refer to DATA_PREPARE.md for the details to prepare the nuScenes and nuScenes-C datasets.

Getting Started

Kindly refer to GET_STARTED.md to learn more usage about this codebase.

Model Zoo

 Camera-Only BEV Detection
 Camera-Only Monocular 3D Detection
 LiDAR-Camera Fusion BEV Detection
 Camera-Only BEV Map Segmentation
 Multi-Camera Depth Estimation
 Multi-Camera Semantic Occupancy Prediction

Robustness Benchmark

📐 Metrics: The nuScenes Detection Score (NDS) is consistently used as the main indicator for evaluating model performance in our benchmark. The following two metrics are adopted to compare between models' robustness:

  • mCE (the lower the better): The average corruption error (in percentage) of a candidate model compared to the baseline model, which is calculated among all corruption types across three severity levels.
  • mRR (the higher the better): The average resilience rate (in percentage) of a candidate model compared to its "clean" performance, which is calculated among all corruption types across three severity levels.

⚙️ Notation: Symbol ⭐ denotes the baseline model adopted in mCE calculation. For more detailed experimental results, please refer to RESULTS.md.

BEV Detection

Model mCE (%) $\downarrow$ mRR (%) $\uparrow$ Clean Cam Crash Frame Lost Color Quant Motion Blur Bright Low Light Fog Snow
DETR3D⭐ 100.00 70.77 0.4224 0.2859 0.2604 0.3177 0.2661 0.4002 0.2786 0.3912 0.1913
DETR3DCBGS 99.21 70.02 0.4341 0.2991 0.2685 0.3235 0.2542 0.4154 0.2766 0.4020 0.1925
BEVFormerSmall 101.23 59.07 0.4787 0.2771 0.2459 0.3275 0.2570 0.3741 0.2413 0.3583 0.1809
BEVFormerBase 97.97 60.40 0.5174 0.3154 0.3017 0.3509 0.2695 0.4184 0.2515 0.4069 0.1857
PETRR50-p4 111.01 61.26 0.3665 0.2320 0.2166 0.2472 0.2299 0.2841 0.1571 0.2876 0.1417
PETRVoV-p4 100.69 65.03 0.4550 0.2924 0.2792 0.2968 0.2490 0.3858 0.2305 0.3703 0.2632
ORA3D 99.17 68.63 0.4436 0.3055 0.2750 0.3360 0.2647 0.4075 0.2613 0.3959 0.1898
BEVDetR50 115.12 51.83 0.3770 0.2486 0.1924 0.2408 0.2061 0.2565 0.1102 0.2461 0.0625
BEVDetR101 113.68 53.12 0.3877 0.2622 0.2065 0.2546 0.2265 0.2554 0.1118 0.2495 0.0810
BEVDetR101-pt 112.80 56.35 0.3780 0.2442 0.1962 0.3041 0.2590 0.2599 0.1398 0.2073 0.0939
BEVDetSwinT 116.48 46.26 0.4037 0.2609 0.2115 0.2278 0.2128 0.2191 0.0490 0.2450 0.0680
BEVDepthR50 110.02 56.82 0.4058 0.2638 0.2141 0.2751 0.2513 0.2879 0.1757 0.2903 0.0863
BEVerseSwinT 110.67 48.60 0.4665 0.3181 0.3037 0.2600 0.2647 0.2656 0.0593 0.2781 0.0644
BEVerseSwinS 117.82 49.57 0.4951 0.3364 0.2485 0.2807 0.2632 0.3394 0.1118 0.2849 0.0985
PolarFormerR101 96.06 70.88 0.4602 0.3133 0.2808 0.3509 0.3221 0.4304 0.2554 0.4262 0.2304
PolarFormerVoV 98.75 67.51 0.4558 0.3135 0.2811 0.3076 0.2344 0.4280 0.2441 0.4061 0.2468
SRCN3DR101 99.67 70.23 0.4286 0.2947 0.2681 0.3318 0.2609 0.4074 0.2590 0.3940 0.1920
SRCN3DVoV 102.04 67.95 0.4205 0.2875 0.2579 0.2827 0.2143 0.3886 0.2274 0.3774 0.2499
Sparse4DR101 100.01 55.04 0.5438 0.2873 0.2611 0.3310 0.2514 0.3984 0.2510 0.3884 0.2259
SOLOFusionshort 108.68 61.45 0.3907 0.2541 0.2195 0.2804 0.2603 0.2966 0.2033 0.2998 0.1066
SOLOFusionlong 97.99 64.42 0.4850 0.3159 0.2490 0.3598 0.3460 0.4002 0.2814 0.3991 0.1480
SOLOFusionfusion 92.86 64.53 0.5381 0.3806 0.3464 0.4058 0.3642 0.4329 0.2626 0.4480 0.1376
FCOS3Dfinetune 107.82 62.09 0.3949 0.2849 0.2479 0.2574 0.2570 0.3218 0.1468 0.3321 0.1136
BEVFusionCam 109.02 57.81 0.4121 0.2777 0.2255 0.2763 0.2788 0.2902 0.1076 0.3041 0.1461
BEVFusionLiDAR - - 0.6928 - - - - - - - -
BEVFusionC+L 43.80 97.41 0.7138 0.6963 0.6931 0.7044 0.6977 0.7018 0.6787 - -
TransFusion - - 0.6887 0.6843 0.6447 0.6819 0.6749 0.6843 0.6663 - -
AutoAlignV2 - - 0.6139 0.5849 0.5832 0.6006 0.5901 0.6076 0.5770 - -

Multi-Camera Depth Estimation

Model Metric Clean Cam Crash Frame Lost Color Quant Motion Blur Bright Low Light Fog Snow
SurroundDepth Abs Rel 0.280 0.485 0.497 0.334 0.338 0.339 0.354 0.320 0.423

Multi-Camera Semantic Occupancy Prediction

Model Metric Clean Cam Crash Frame Lost Color Quant Motion Blur Bright Low Light Fog Snow
TPVFormer mIoU vox 52.06 27.39 22.85 38.16 38.64 49.00 37.38 46.69 19.39
SurroundOcc SC mIoU 20.30 11.60 10.00 14.03 12.41 19.18 12.15 18.42 7.39

BEV Model Calibration

Model Pretrain Temporal Depth CBGS Backbone EncoderBEV Input Size mCE (%) mRR (%) NDS
DETR3D ✓ ✗ ✗ ✗ ResNet Attention 1600×900 100.00 70.77 0.4224
DETR3DCBGS ✓ ✗ ✗ ✓ ResNet Attention 1600×900 99.21 70.02 0.4341
BEVFormerSmall ✓ ✓ ✗ ✗ ResNet Attention 1280×720 101.23 59.07 0.4787
BEVFormerBase ✓ ✓ ✗ ✗ ResNet Attention 1600×900 97.97 60.40 0.5174
PETRR50-p4 ✗ ✗ ✗ ✗ ResNet Attention 1408×512 111.01 61.26 0.3665
PETRVoV-p4 ✓ ✗ ✗ ✗ VoVNetV2 Attention 1600×900 100.69 65.03 0.4550
ORA3D ✓ ✗ ✗ ✗ ResNet Attention 1600×900 99.17 68.63 0.4436
PolarFormerR101 ✓ ✗ ✗ ✗ ResNet Attention 1600×900 96.06 70.88 0.4602
PolarFormerVoV ✓ ✗ ✗ ✗ VoVNetV2 Attention 1600×900 98.75 67.51 0.4558
SRCN3DR101 ✓ ✗ ✗ ✗ ResNet CNN+Attn. 1600×900 99.67 70.23 0.4286
SRCN3DVoV ✓ ✗ ✗ ✗ VoVNetV2 CNN+Attn. 1600×900 102.04 67.95 0.4205
Sparse4DR101 ✓ ✓ ✗ ✗ ResNet CNN+Attn. 1600×900 100.01 55.04 0.5438
BEVDetR50 ✗ ✗ ✓ ✓ ResNet CNN 704×256 115.12 51.83 0.3770
BEVDetR101 ✗ ✗ ✓ ✓ ResNet CNN 704×256 113.68 53.12 0.3877
BEVDetR101-pt ✓ ✗ ✓ ✓ ResNet CNN 704×256 112.80 56.35 0.3780
BEVDetSwinT ✗ ✗ ✓ ✓ Swin CNN 704×256 116.48 46.26 0.4037
BEVDepthR50 ✗ ✗ ✓ ✓ ResNet CNN 704×256 110.02 56.82 0.4058
BEVerseSwinT ✗ ✗ ✓ ✓ Swin CNN 704×256 137.25 28.24 0.1603
BEVerseSwinT ✗ ✓ ✓ ✓ Swin CNN 704×256 110.67 48.60 0.4665
BEVerseSwinS ✗ ✗ ✓ ✓ Swin CNN 1408×512 132.13 29.54 0.2682
BEVerseSwinS ✗ ✓ ✓ ✓ Swin CNN 1408×512 117.82 49.57 0.4951
SOLOFusionshort ✗ ✓ ✓ ✗ ResNet CNN 704×256 108.68 61.45 0.3907
SOLOFusionlong ✗ ✓ ✓ ✗ ResNet CNN 704×256 97.99 64.42 0.4850
SOLOFusionfusion ✗ ✓ ✓ ✓ ResNet CNN 704×256 92.86 64.53 0.5381

Note: Pretrain denotes models initialized from the FCOS3D checkpoint. Temporal indicates whether temporal information is used. Depth denotes models with an explicit depth estimation branch. CBGS highlight models use the class-balanced group-sampling strategy.

Create Corruption Set

You can manage to create your own "RoboBEV" corrpution sets! Follow the instructions listed in CREATE.md.

TODO List

  • Initial release. 🚀
  • Add scripts for creating common corruptions.
  • Add download link of nuScenes-C.
  • Add evaluation scripts on corruption sets.
  • Establish benchmark for BEV map segmentation.
  • Establish benchmark for multi-camera depth estimation.
  • Establish benchmark for multi-camera semantic occupancy prediction.
  • ...

Citation

If you find this work helpful, please kindly consider citing the following:

@article{xie2024benchmarking,
    title = {Benchmarking and Improving Bird's Eye View Perception Robustness in Autonomous Driving},
    author = {Xie, Shaoyuan and Kong, Lingdong and Zhang, Wenwei and Ren, Jiawei and Pan, Liang and Chen, Kai and Liu, Ziwei},
    journal = {arXiv preprint arXiv:2405.17426}, 
    year = {2024}
}
@article{xie2023robobev,
    title = {RoboBEV: Towards Robust Bird's Eye View Perception under Corruptions},
    author = {Xie, Shaoyuan and Kong, Lingdong and Zhang, Wenwei and Ren, Jiawei and Pan, Liang and Chen, Kai and Liu, Ziwei},
    journal = {arXiv preprint arXiv:2304.06719}, 
    year = {2023}
}
@misc{xie2023robobev_codebase,
    title = {The RoboBEV Benchmark for Robust Bird's Eye View Detection under Common Corruption and Domain Shift},
    author = {Xie, Shaoyuan and Kong, Lingdong and Zhang, Wenwei and Ren, Jiawei and Pan, Liang and Chen, Kai and Liu, Ziwei},
    howpublished = {\url{https://github.com/Daniel-xsy/RoboBEV}},
    year = {2023}
}

License

Creative Commons License
This work is under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License, while some specific operations in this codebase might be with other licenses. Please refer to LICENSE.md for a more careful check, if you are using our code for commercial matters.

Acknowledgements

This work is developed based on the MMDetection3D codebase.


MMDetection3D is an open source object detection toolbox based on PyTorch, towards the next-generation platform for general 3D detection. It is a part of the OpenMMLab project developed by MMLab.

❤️ We thank Jiangmiao Pang and Tai Wang for their insightful discussions and feedback. We thank the OpenDataLab platform for hosting our datasets.