English | 简体中文

Robo3D: Towards Robust and Reliable 3D Perception against Corruptions

Lingdong Kong^1,2,*    Youquan Liu^1,3,*    Xin Li^1,4,*    Runnan Chen^1,5    Wenwei Zhang^1,6
Jiawei Ren⁶    Liang Pan⁶    Kai Chen¹    Ziwei Liu^6
1Shanghai AI Laboratory    ²National University of Singapore    ³Hochschule Bremerhaven    ⁴East China Normal University    ⁵The University of Hong Kong    ⁶S-Lab, Nanyang Technological University

About

Robo3D is an evaluation suite heading toward robust and reliable 3D perception in autonomous driving. With it, we probe the robustness of 3D detectors and segmentors under out-of-distribution (OoD) scenarios against corruptions that occur in the real-world environment. Specifically, we consider natural corruptions happen in the following cases:

1. Adverse weather conditions, such as fog, wet ground, and snow;
2. External disturbances that are caused by motion blur or result in LiDAR beam missing;
3. Internal sensor failure, including crosstalk, possible incomplete echo, and cross-sensor scenarios.

Clean	Fog	Wet Ground
Snow	Motion Blur	Beam Missing
Crosstalk	Incomplete Echo	Cross-Sensor

Visit our project page to explore more examples. 🚘

Updates

• [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. 🚙
• [2023.09] - Intend to improve the OoD robustness of your 3D perception models? Check out our recent work, Seal 🦭, an image-to-LiDAR self-supervised pretraining framework that leverages off-the-shelf knowledge from vision foundation models for cross-modality representation learning.
• [2023.07] - Robo3D was accepted to ICCV 2023! 🎉
• [2023.03] - We establish "Robust 3D Perception" leaderboards on Paper-with-Code: ¹KITTI-C, ²SemanticKITTI-C, ³nuScenes-C, and ⁴WOD-C. Join the challenge today! 🙋
• [2023.03] - The KITTI-C, SemanticKITTI-C, and nuScenes-C datasets are ready for download at the OpenDataLab platform. Kindly refer to this page for more details on preparing these datasets. 🍻
• [2023.01] - Launch of the Robo3D benchmark. In this initial version, we include 12 detectors and 22 segmentors, evaluated on 4 large-scale autonomous driving datasets (KITTI, SemanticKITTI, nuScenes, and Waymo Open) with 8 corruption types across 3 severity levels.

Outline

• Taxonomy
• Video Demo
• Installation
• Data Preparation
• Getting Started
• Model Zoo
• Benchmark
• Create Corruption Set
• TODO List
• Citation
• License
• Acknowledgements

Taxonomy

Fog	Wet Ground	Snow	Motion Blur
Beam Missing	Crosstalk	Incomplete Echo	Cross-Sensor

Video Demo

Demo 1	Demo 2	Demo 3
Link ⤴️	Link ⤴️	Link ⤴️

Installation

For details related to installation, kindly refer to INSTALL.md.

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 ¹KITTI, ²KITTI-C, ³SemanticKITTI, ⁴SemanticKITTI-C, ⁵nuScenes, ⁶nuScenes-C, ⁷WOD, and ⁸WOD-C datasets.

Getting Started

To learn more usage about this codebase, kindly refer to GET_STARTED.md.

Model Zoo

 LiDAR Semantic Segmentation

• SqueezeSeg, ICRA 2018. ^[Code]
• SqueezeSegV2, ICRA 2019. ^[Code]
• MinkowskiNet, CVPR 2019. ^[Code]
• RangeNet++, IROS 2019. ^[Code]
• KPConv, ICCV 2019. ^[Code]
• SalsaNext, ISVC 2020. ^[Code]
• RandLA-Net, CVPR 2020. ^[Code]
• PolarNet, CVPR 2020. ^[Code]
• 3D-MiniNet, IROS 2020. ^[Code]
• SPVCNN, ECCV 2020. ^[Code]
• Cylinder3D, CVPR 2021. ^[Code]
• FIDNet, IROS 2021. ^[Code]
• RPVNet, ICCV 2021.
• CENet, ICME 2022. ^[Code]
• CPGNet, ICRA 2022. ^[Code]
• 2DPASS, ECCV 2022. ^[Code]
• GFNet, TMLR 2022. ^[Code]
• PCB-RandNet, arXiv 2022. ^[Code]
• PIDS, WACV 2023. ^[Code]
• SphereFormer, CVPR 2023. ^[Code]
• WaffleIron, ICCV 2023. ^[Code]
• FRNet, arXiv 2023. ^[Code]

 LiDAR Panoptic Segmentation

• DS-Net, CVPR 2021. ^[Code]
• Panoptic-PolarNet, CVPR 2021. ^[Code]

 3D Object Detection

• SECOND, Sensors 2018. ^[Code]
• PointPillars, CVPR 2019. ^[Code]
• PointRCNN, CVPR 2019. ^[Code]
• Part-A2, T-PAMI 2020.
• PV-RCNN, CVPR 2020. ^[Code]
• 3DSSD, CVPR 2020. ^[Code]
• SA-SSD, CVPR 2020. ^[Code]
• CenterPoint, CVPR 2021. ^[Code]
• PV-RCNN++, IJCV 2022. ^[Code]
• SphereFormer, CVPR 2023. ^[Code]

Benchmark

LiDAR Semantic Segmentation

The mean Intersection-over-Union (mIoU) is consistently used as the main indicator for evaluating model performance in our LiDAR semantic segmentation benchmark. The following two metrics are adopted to compare among 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.

🚗  SemanticKITTI-C

Model	mCE (%)	mRR (%)	Clean	Fog	Wet Ground	Snow	Motion Blur	Beam Missing	Cross-Talk	Incomplete Echo	Cross-Sensor
SqueezeSeg	164.87	66.81	31.61	18.85	27.30	22.70	17.93	25.01	21.65	27.66	7.85
SqueezeSegV2	152.45	65.29	41.28	25.64	35.02	27.75	22.75	32.19	26.68	33.80	11.78
RangeNet21	136.33	73.42	47.15	31.04	40.88	37.43	31.16	38.16	37.98	41.54	18.76
RangeNet53	130.66	73.59	50.29	36.33	43.07	40.02	30.10	40.80	46.08	42.67	16.98
SalsaNext	116.14	80.51	55.80	34.89	48.44	45.55	47.93	49.63	40.21	48.03	44.72
FIDNet34	113.81	76.99	58.80	43.66	51.63	49.68	40.38	49.32	49.46	48.17	29.85
CENet34	103.41	81.29	62.55	42.70	57.34	53.64	52.71	55.78	45.37	53.40	45.84
FRNet	96.80	80.04	67.55	47.61	62.15	57.08	56.80	62.54	40.94	58.11	47.30
KPConv	99.54	82.90	62.17	54.46	57.70	54.15	25.70	57.35	53.38	55.64	53.91
PIDSNAS1.25x	104.13	77.94	63.25	47.90	54.48	48.86	22.97	54.93	56.70	55.81	52.72
PIDSNAS2.0x	101.20	78.42	64.55	51.19	55.97	51.11	22.49	56.95	57.41	55.55	54.27
WaffleIron	109.54	72.18	66.04	45.52	58.55	49.30	33.02	59.28	22.48	58.55	54.62
PolarNet	118.56	74.98	58.17	38.74	50.73	49.42	41.77	54.10	25.79	48.96	39.44
⭐MinkUNet18	100.00	81.90	62.76	55.87	53.99	53.28	32.92	56.32	58.34	54.43	46.05
MinkUNet34	100.61	80.22	63.78	53.54	54.27	50.17	33.80	57.35	58.38	54.88	46.95
Cylinder3DSPC	103.25	80.08	63.42	37.10	57.45	46.94	52.45	57.64	55.98	52.51	46.22
Cylinder3DTSC	103.13	83.90	61.00	37.11	53.40	45.39	58.64	56.81	53.59	54.88	49.62
SPVCNN18	100.30	82.15	62.47	55.32	53.98	51.42	34.53	56.67	58.10	54.60	45.95
SPVCNN34	99.16	82.01	63.22	56.53	53.68	52.35	34.39	56.76	59.00	54.97	47.07
RPVNet	111.74	73.86	63.75	47.64	53.54	51.13	47.29	53.51	22.64	54.79	46.17
CPGNet	107.34	81.05	61.50	37.79	57.39	51.26	59.05	60.29	18.50	56.72	57.79
2DPASS	106.14	77.50	64.61	40.46	60.68	48.53	57.80	58.78	28.46	55.84	50.01
GFNet	108.68	77.92	63.00	42.04	56.57	56.71	58.59	56.95	17.14	55.23	49.48

Note: Symbol ^⭐ denotes the baseline model adopted in mCE calculation.

🚙  nuScenes-C

Model	mCE (%)	mRR (%)	Clean	Fog	Wet Ground	Snow	Motion Blur	Beam Missing	Cross-Talk	Incomplete Echo	Cross-Sensor
FIDNet34	122.42	73.33	71.38	64.80	68.02	58.97	48.90	48.14	57.45	48.76	23.70
CENet34	112.79	76.04	73.28	67.01	69.87	61.64	58.31	49.97	60.89	53.31	24.78
FRNet	98.63	77.48	77.65	69.14	76.58	69.49	54.49	68.32	41.43	58.74	43.13
WaffleIron	106.73	72.78	76.07	56.07	73.93	49.59	59.46	65.19	33.12	61.51	44.01
PolarNet	115.09	76.34	71.37	58.23	69.91	64.82	44.60	61.91	40.77	53.64	42.01
⭐MinkUNet18	100.00	74.44	75.76	53.64	73.91	40.35	73.39	68.54	26.58	63.83	50.95
MinkUNet34	96.37	75.08	76.90	56.91	74.93	37.50	75.24	70.10	29.32	64.96	52.96
Cylinder3DSPC	111.84	72.94	76.15	59.85	72.69	58.07	42.13	64.45	44.44	60.50	42.23
Cylinder3DTSC	105.56	78.08	73.54	61.42	71.02	58.40	56.02	64.15	45.36	59.97	43.03
SPVCNN18	106.65	74.70	74.40	59.01	72.46	41.08	58.36	65.36	36.83	62.29	49.21
SPVCNN34	97.45	75.10	76.57	55.86	74.04	41.95	74.63	68.94	28.11	64.96	51.57
2DPASS	98.56	75.24	77.92	64.50	76.76	54.46	62.04	67.84	34.37	63.19	45.83
GFNet	92.55	83.31	76.79	69.59	75.52	71.83	59.43	64.47	66.78	61.86	42.30

Note: Symbol ^⭐ denotes the baseline model adopted in mCE calculation.

🚕  WOD-C

Model	mCE (%)	mRR (%)	Clean	Fog	Wet Ground	Snow	Motion Blur	Beam Missing	Cross-Talk	Incomplete Echo	Cross-Sensor
⭐MinkUNet18	100.00	91.22	69.06	66.99	60.99	57.75	68.92	64.15	65.37	63.36	56.44
MinkUNet34	96.21	91.80	70.15	68.31	62.98	57.95	70.10	65.79	66.48	64.55	59.02
Cylinder3DTSC	106.02	92.39	65.93	63.09	59.40	58.43	65.72	62.08	62.99	60.34	55.27
SPVCNN18	103.60	91.60	67.35	65.13	59.12	58.10	67.24	62.41	65.46	61.79	54.30
SPVCNN34	98.72	92.04	69.01	67.10	62.41	57.57	68.92	64.67	64.70	64.14	58.63

Note: Symbol ^⭐ denotes the baseline model adopted in mCE calculation.

3D Object Detection

The mean average precision (mAP) and nuScenes detection score (NDS) are consistently used as the main indicator for evaluating model performance in our LiDAR semantic segmentation 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.

🚗  KITTI-C

Model	mCE (%)	mRR (%)	Clean	Fog	Wet Ground	Snow	Motion Blur	Beam Missing	Cross-Talk	Incomplete Echo	Cross-Sensor
PointPillars	110.67	74.94	66.70	45.70	66.71	35.77	47.09	52.24	60.01	54.84	37.50
SECOND	95.93	82.94	68.49	53.24	68.51	54.92	49.19	54.14	67.19	59.25	48.00
PointRCNN	91.88	83.46	70.26	56.31	71.82	50.20	51.52	56.84	65.70	62.02	54.73
PartA2Free	82.22	81.87	76.28	58.06	76.29	58.17	55.15	59.46	75.59	65.66	51.22
PartA2Anchor	88.62	80.67	73.98	56.59	73.97	51.32	55.04	56.38	71.72	63.29	49.15
PVRCNN	90.04	81.73	72.36	55.36	72.89	52.12	54.44	56.88	70.39	63.00	48.01
⭐CenterPoint	100.00	79.73	68.70	53.10	68.71	48.56	47.94	49.88	66.00	58.90	45.12
SphereFormer	-	-	-	-	-	-	-	-	-	-	-

Note: Symbol ^⭐ denotes the baseline model adopted in mCE calculation.

🚙  nuScenes-C

Model	mCE (%)	mRR (%)	Clean	Fog	Wet Ground	Snow	Motion Blur	Beam Missing	Cross-Talk	Incomplete Echo	Cross-Sensor
PointPillarsMH	102.90	77.24	43.33	33.16	42.92	29.49	38.04	33.61	34.61	30.90	25.00
SECONDMH	97.50	76.96	47.87	38.00	47.59	33.92	41.32	35.64	40.30	34.12	23.82
⭐CenterPoint	100.00	76.68	45.99	35.01	45.41	31.23	41.79	35.16	35.22	32.53	25.78
CenterPointLR	98.74	72.49	49.72	36.39	47.34	32.81	40.54	34.47	38.11	35.50	23.16
CenterPointHR	95.80	75.26	50.31	39.55	49.77	34.73	43.21	36.21	40.98	35.09	23.38
SphereFormer	-	-	-	-	-	-	-	-	-	-	-

Note: Symbol ^⭐ denotes the baseline model adopted in mCE calculation.

🚕  WOD-C

Model	mCE (%)	mRR (%)	Clean	Fog	Wet Ground	Snow	Motion Blur	Beam Missing	Cross-Talk	Incomplete Echo	Cross-Sensor
PointPillars	127.53	81.23	50.17	31.24	49.75	46.07	34.93	43.93	39.80	43.41	36.67
SECOND	121.43	81.12	53.37	32.89	52.99	47.20	35.98	44.72	49.28	46.84	36.43
PVRCNN	104.90	82.43	61.27	37.32	61.27	60.38	42.78	49.53	59.59	54.43	38.73
⭐CenterPoint	100.00	83.30	63.59	43.06	62.84	58.59	43.53	54.41	60.32	57.01	43.98
PVRCNN++	91.60	84.14	67.45	45.50	67.18	62.71	47.35	57.83	64.71	60.96	47.77
SphereFormer	-	-	-	-	-	-	-	-	-	-	-

Note: Symbol ^⭐ denotes the baseline model adopted in mCE calculation.

🚦 More Benchmarking Results

For more detailed experimental results and visual comparisons, please refer to RESULTS.md.

Create Corruption Set

You can manage to create your own "Robo3D" corruption sets on other LiDAR-based point cloud datasets using our defined corruption types! Follow the instructions listed in CREATE.md.

TODO List

• Initial release. 🚀
• Add scripts for creating common corruptions.
• Add download links for corruption sets.
• Add evaluation scripts on corruption sets.
• Release checkpoints.
• ...

Citation

If you find this work helpful, please kindly consider citing our paper:

@inproceedings{kong2023robo3d,
    author = {Lingdong Kong and Youquan Liu and Xin Li and Runnan Chen and Wenwei Zhang and Jiawei Ren and Liang Pan and Kai Chen and Ziwei Liu},
    title = {Robo3D: Towards Robust and Reliable 3D Perception against Corruptions},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    pages = {19994--20006},
    year = {2023},
}

@misc{kong2023robo3d_benchmark,
  title = {The Robo3D Benchmark for Robust and Reliable 3D Perception},
  author = {Lingdong Kong and Youquan Liu and Xin Li and Runnan Chen and Wenwei Zhang and Jiawei Ren and Liang Pan and Kai Chen and Ziwei Liu},
  howpublished = {\url{https://github.com/ldkong1205/Robo3D}},
  year = {2023},
}

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.