English | 简体中文

RoboDepth: Robust Out-of-Distribution Depth Estimation under Corruptions

Lingdong Kong^1,2   Shaoyuan Xie³   Hanjiang Hu⁴   Lai Xing Ng^2,5   Benoit R. Cottereau^2,6   Wei Tsang Ooi^1,2
1National University of Singapore    ²CNRS@CREATE    ³University of California, Irvine    ⁴Carnegie Mellon University    ⁵Institute for Infocomm Research, A*STAR    ⁶CNRS

About

RoboDepth is a comprehensive evaluation benchmark designed for probing the robustness of monocular depth estimation algorithms. It includes 18 common corruption types, ranging from weather and lighting conditions, sensor failure and movement, and noises during data processing.

Updates

• [2024.01] - The toolkit tailored for the RoboDrive Challenge has been released. 🛠️
• [2023.12] - We are hosting the RoboDrive Challenge at ICRA 2024. 🚙
• [2023.09] - RoboDepth was accepted to NeurIPS 2023 Track on Datasets and Benchmarks! 🎉
• [2023.08] - We support robust depth estimation on real-world scenarios, including nuScenes, nuScenes-Night, Cityscapes, and Foggy-Cityscapes. See here for more details.
• [2023.08] - We establish the nuScenes-C benchmark for robust multi-view depth estimation. See here for more details.
• [2023.07] - The technical report of the RoboDepth Challenge is available on arXiv.
• [2023.06] - We have successfully concluded the RoboDepth Challenge! Key statistics of this competition: 226 teams registered at CodaLab, 66 of which made a total number of 1137 valid submissions. More details are included in these slides. We thank the exceptional support from our participants! ❤️
• [2023.06] - We are glad to announce the winning teams of this competition:
  • Track 1: 🥇 OpenSpaceAI, 🥈 USTC-IAT-United, 🥉 YYQ.
  • Track 2: 🥇 USTCxNetEaseFuxi, 🥈 OpenSpaceAI, 🥉 GANCV.
  • Innovation Prize: 🎖️ Scent-Depth, 🎖️ Ensemble, 🎖️ AIIA-RDepth.
• [2023.06] - The video recordings of the RoboDepth Workshop are out. Know more details about how our participants were dedicated to improving the robustness of depth estimation models. 🎥
• [2023.05] - Glad to announce that the RoboDepth Challenge will be sponsored by Baidu Research. 🍻
• [2023.01] - The NYUDepth2-C dataset is ready to be downloaded! See here for more details.
• [2023.01] - Evaluation server for Track 2 (fully-supervised depth estimation) is available on this page.
• [2023.01] - Evaluation server for Track 1 (self-supervised depth estimation) is available on this page.
• [2022.11] - We are organizing the RoboDepth Challenge at ICRA 2023. Join the challenge today! 🙋
• [2022.11] - The KITTI-C dataset is ready to be downloaded! See here for more details.

Outline

• Installation
• Data Preparation
• Getting Started
• Model Zoo
• Benchmark
• Idiosyncrasy Analysis
• Create Corruption Sets
• TODO List
• Citation
• License
• Acknowledgements

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.

The RoboDepth Benchmark

Kindly refer to DATA_PREPARE.md for the details to prepare the ¹KITTI, ²KITTI-C, ³NYUDepth2, ⁴NYUDepth2-C, ⁵Cityscapes, ⁶Foggy-Cityscapes, ⁷nuScenes, and ⁸nuScenes-C, datasets.

Competition @ ICRA 2023

Kindly refer to this page for the details to prepare the training and evaluation data associated with the 1st RoboDepth Competition at the 40th IEEE Conference on Robotics and Automation (ICRA 2023).

Getting Started

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

Model Zoo

🚘 - Outdoor Depth Estimation

 Self-Supervised Depth Estimation

• MonoDepth2, ICCV 2019. ^[Code]
• DepthHints, ICCV 2019. ^[Code]
• MaskOcc, arXiv 2019. ^[Code]
• DNet, IROS 2020. ^[Code]
• SGDepth, ECCV 2020. ^[Code]
• CADepth, 3DV 2021. ^[Code]
• TC-Depth, 3DV 2021. ^[Code]
• HR-Depth, AAAI 2021. ^[Code]
• Insta-DM, AAAI 2021. ^[Code]
• DIFFNet, BMVC 2021. ^[Code]
• ManyDepth, CVPR 2021. ^[Code]
• EPCDepth, ICCV 2021. ^[Code]
• FSRE-Depth, ICCV 2021. ^[Code]
• R-MSFM, ICCV 2021. ^[Code]
• MonoViT, 3DV 2022. ^[Code]
• DepthFormer, CVPR 2022. ^[Code]
• DynaDepth, ECCV 2022. ^[Code]
• DynamicDepth, ECCV 2022. ^[Code]
• RA-Depth, ECCV 2022. ^[Code]
• Dyna-DM, arXiv 2022. ^[Code]
• TriDepth, WACV 2023. ^[Code]
• FreqAwareDepth, WACV 2023. ^[Code]
• Lite-Mono, CVPR 2023. ^[Code]

 Self-Supervised Multi-View Depth Estimation

• MonoDepth2, ICCV 2019. ^[Code]
• SurroundDepth, CoRL 2022. ^[Code]

 Fully-Supervised Depth Estimation

• AdaBins, CVPR 2021. ^[Code]
• NeWCRFs, CVPR 2022. ^[Code]
• DepthFormer, arXiv 2022. ^[Code]
• GLPDepth, arXiv 2022. ^[Code]

 Semi-Supervised Depth Estimation

• MaskingDepth, arXiv 2022. ^[Code]

🏠 - Indoor Depth Estimation

 Self-Supervised Depth Estimation

• P²Net, ECCV 2020. ^[Code]
• EPCDepth, ICCV 2021. ^[Code]

 Fully-Supervised Depth Estimation

• BTS, arXiv 2019. ^[Code]
• AdaBins, CVPR 2021. ^[Code]
• DPT, ICCV 2021. ^[Code]
• SimIPU, AAAI 2022. ^[Code]
• NeWCRFs, CVPR 2022. ^[Code]
• P3Depth, CVPR 2022. ^[Code]
• DepthFormer, arXiv 2022. ^[Code]
• GLPDepth, arXiv 2022. ^[Code]
• BinsFormer, arXiv 2022. ^[Code]

 Semi-Supervised Depth Estimation

• MaskingDepth, arXiv 2022. ^[Code]

Benchmark

📊 Metrics: The following metrics are consistently used in our benchmark:

• Absolute Relative Difference (the lower the better): $\text{Abs Rel} = \frac{1}{|D|}\sum_{pred\in D}\frac{|gt - pred|}{gt}$ .

• Accuracy (the higher the better): $\delta_t = \frac{1}{|D|}|{\ pred\in D | \max{(\frac{gt}{pred}, \frac{pred}{gt})< 1.25^t}}| \times 100\%$ .

• Depth Estimation Error (the lower the better):

  • $\text{DEE}_1 = \text{Abs Rel} - \delta_1 + 1$ ;
  • $\text{DEE}_2 = \frac{\text{Abs Rel} - \delta_1 + 1}{2}$ ;
  • $\text{DEE}_3 = \frac{\text{Abs Rel}}{\delta_1}$ .

• The second Depth Estimation Error term ($\text{DEE}_2$) is adopted as the main indicator for evaluating model performance in our RoboDepth 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, which is calculated among all corruption types across five 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 five severity levels.

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

KITTI-C

Model	Modality	mCE (%)	mRR (%)	Clean	Bright	Dark	Fog	Frost	Snow	Contrast	Defocus	Glass	Motion	Zoom	Elastic	Quant	Gaussian	Impulse	Shot	ISO	Pixelate	JPEG
MonoDepth2R18⭐	Mono	100.00	84.46	0.119	0.130	0.280	0.155	0.277	0.511	0.187	0.244	0.242	0.216	0.201	0.129	0.193	0.384	0.389	0.340	0.388	0.145	0.196
MonoDepth2R18+nopt	Mono	119.75	82.50	0.144	0.183	0.343	0.311	0.312	0.399	0.416	0.254	0.232	0.199	0.207	0.148	0.212	0.441	0.452	0.402	0.453	0.153	0.171
MonoDepth2R18+HR	Mono	106.06	82.44	0.114	0.129	0.376	0.155	0.271	0.582	0.214	0.393	0.257	0.230	0.232	0.123	0.215	0.326	0.352	0.317	0.344	0.138	0.198
MonoDepth2R50	Mono	113.43	80.59	0.117	0.127	0.294	0.155	0.287	0.492	0.233	0.427	0.392	0.277	0.208	0.130	0.198	0.409	0.403	0.368	0.425	0.155	0.211
MaskOcc	Mono	104.05	82.97	0.117	0.130	0.285	0.154	0.283	0.492	0.200	0.318	0.295	0.228	0.201	0.129	0.184	0.403	0.410	0.364	0.417	0.143	0.177
DNetR18	Mono	104.71	83.34	0.118	0.128	0.264	0.156	0.317	0.504	0.209	0.348	0.320	0.242	0.215	0.131	0.189	0.362	0.366	0.326	0.357	0.145	0.190
CADepth	Mono	110.11	80.07	0.108	0.121	0.300	0.142	0.324	0.529	0.193	0.356	0.347	0.285	0.208	0.121	0.192	0.423	0.433	0.383	0.448	0.144	0.195
HR-Depth	Mono	103.73	82.93	0.112	0.121	0.289	0.151	0.279	0.481	0.213	0.356	0.300	0.263	0.224	0.124	0.187	0.363	0.373	0.336	0.374	0.135	0.176
DIFFNetHRNet	Mono	94.96	85.41	0.102	0.111	0.222	0.131	0.199	0.352	0.161	0.513	0.330	0.280	0.197	0.114	0.165	0.292	0.266	0.255	0.270	0.135	0.202
ManyDepthsingle	Mono	105.41	83.11	0.123	0.135	0.274	0.169	0.288	0.479	0.227	0.254	0.279	0.211	0.194	0.134	0.189	0.430	0.450	0.387	0.452	0.147	0.182
FSRE-Depth	Mono	99.05	83.86	0.109	0.128	0.261	0.139	0.237	0.393	0.170	0.291	0.273	0.214	0.185	0.119	0.179	0.400	0.414	0.370	0.407	0.147	0.224
MonoViTMPViT	Mono	79.33	89.15	0.099	0.106	0.243	0.116	0.213	0.275	0.119	0.180	0.204	0.163	0.179	0.118	0.146	0.310	0.293	0.271	0.290	0.162	0.154
MonoViTMPViT+HR	Mono	74.95	89.72	0.094	0.102	0.238	0.114	0.225	0.269	0.117	0.145	0.171	0.145	0.184	0.108	0.145	0.302	0.277	0.259	0.285	0.135	0.148
DynaDepthR18	Mono	110.38	81.50	0.117	0.128	0.289	0.156	0.289	0.509	0.208	0.501	0.347	0.305	0.207	0.127	0.186	0.379	0.379	0.336	0.379	0.141	0.180
DynaDepthR50	Mono	119.99	77.98	0.113	0.128	0.298	0.152	0.324	0.549	0.201	0.532	0.454	0.318	0.218	0.125	0.197	0.418	0.437	0.382	0.448	0.153	0.216
RA-DepthHRNet	Mono	112.73	78.79	0.096	0.113	0.314	0.127	0.239	0.413	0.165	0.499	0.368	0.378	0.214	0.122	0.178	0.423	0.403	0.402	0.455	0.175	0.192
TriDepthsingle	Mono	109.26	81.56	0.117	0.131	0.300	0.188	0.338	0.498	0.265	0.268	0.301	0.212	0.190	0.126	0.199	0.418	0.438	0.380	0.438	0.142	0.205
Lite-MonoTiny	Mono	92.92	86.69	0.115	0.127	0.257	0.157	0.225	0.354	0.191	0.257	0.248	0.198	0.186	0.127	0.159	0.358	0.342	0.336	0.360	0.147	0.161
Lite-MonoTiny+HR	Mono	86.71	87.63	0.106	0.119	0.227	0.139	0.282	0.370	0.166	0.216	0.201	0.190	0.202	0.116	0.146	0.320	0.291	0.286	0.312	0.148	0.167
Lite-MonoSmall	Mono	100.34	84.67	0.115	0.127	0.251	0.162	0.251	0.430	0.238	0.353	0.282	0.246	0.204	0.128	0.161	0.350	0.336	0.319	0.356	0.154	0.164
Lite-MonoSmall+HR	Mono	89.90	86.05	0.105	0.119	0.263	0.139	0.263	0.436	0.167	0.188	0.181	0.193	0.214	0.117	0.147	0.366	0.354	0.327	0.355	0.152	0.157
Lite-MonoBase	Mono	93.16	85.99	0.110	0.119	0.259	0.144	0.245	0.384	0.177	0.224	0.237	0.221	0.196	0.129	0.175	0.361	0.340	0.334	0.363	0.151	0.165
Lite-MonoBase+HR	Mono	89.85	85.80	0.103	0.115	0.256	0.135	0.258	0.486	0.164	0.220	0.194	0.213	0.205	0.114	0.154	0.340	0.327	0.321	0.344	0.145	0.156
Lite-MonoLarge	Mono	90.75	85.54	0.102	0.110	0.227	0.126	0.255	0.433	0.149	0.222	0.225	0.220	0.192	0.121	0.148	0.363	0.348	0.329	0.362	0.160	0.184
Lite-MonoLarge+HR	Mono	92.01	83.90	0.096	0.112	0.241	0.122	0.280	0.482	0.141	0.193	0.194	0.213	0.222	0.108	0.140	0.403	0.404	0.365	0.407	0.139	0.182
MonoDepth2R18	Stereo	117.69	79.05	0.123	0.133	0.348	0.161	0.305	0.515	0.234	0.390	0.332	0.264	0.209	0.135	0.200	0.492	0.509	0.463	0.493	0.144	0.194
MonoDepth2R18+nopt	Stereo	128.98	79.20	0.150	0.181	0.422	0.292	0.352	0.435	0.342	0.266	0.232	0.217	0.229	0.156	0.236	0.539	0.564	0.521	0.556	0.164	0.178
MonoDepth2R18+HR	Stereo	111.46	81.65	0.117	0.132	0.285	0.167	0.356	0.529	0.238	0.432	0.312	0.279	0.246	0.130	0.206	0.343	0.343	0.322	0.344	0.150	0.209
DepthHints	Stereo	111.41	80.08	0.113	0.124	0.310	0.137	0.321	0.515	0.164	0.350	0.410	0.263	0.196	0.130	0.192	0.440	0.447	0.412	0.455	0.157	0.192
DepthHintsHR	Stereo	112.02	79.53	0.104	0.122	0.282	0.141	0.317	0.480	0.180	0.459	0.363	0.320	0.262	0.118	0.183	0.397	0.421	0.380	0.424	0.141	0.183
DepthHintsHR+nopt	Stereo	141.61	73.18	0.134	0.173	0.476	0.301	0.374	0.463	0.393	0.357	0.289	0.241	0.231	0.142	0.247	0.613	0.658	0.599	0.692	0.152	0.191
MonoDepth2R18	M+S	124.31	75.36	0.116	0.127	0.404	0.150	0.295	0.536	0.199	0.447	0.346	0.283	0.204	0.128	0.203	0.577	0.605	0.561	0.629	0.136	0.179
MonoDepth2R18+nopt	M+S	136.25	76.72	0.146	0.193	0.460	0.328	0.421	0.428	0.440	0.228	0.221	0.216	0.230	0.153	0.229	0.570	0.596	0.549	0.606	0.161	0.177
MonoDepth2R18+HR	M+S	106.06	82.44	0.114	0.129	0.376	0.155	0.271	0.582	0.214	0.393	0.257	0.230	0.232	0.123	0.215	0.326	0.352	0.317	0.344	0.138	0.198
CADepth	M+S	118.29	76.68	0.110	0.123	0.357	0.137	0.311	0.556	0.169	0.338	0.412	0.260	0.193	0.126	0.186	0.546	0.559	0.524	0.582	0.145	0.192
MonoViTMPViT	M+S	75.39	90.39	0.098	0.104	0.245	0.122	0.213	0.215	0.131	0.179	0.184	0.161	0.168	0.112	0.147	0.277	0.257	0.242	0.260	0.147	0.144
MonoViTMPViT+HR	M+S	70.79	90.67	0.090	0.097	0.221	0.113	0.217	0.253	0.113	0.146	0.159	0.144	0.175	0.098	0.138	0.267	0.246	0.236	0.246	0.135	0.145

NYUDepth2-C

Model	mCE (%)	mRR (%)	Clean	Bright	Dark	Contrast	Defocus	Glass	Motion	Zoom	Elastic	Quant	Gaussian	Impulse	Shot	ISO	Pixelate	JPEG
BTSR50	122.78	80.63	0.122	0.149	0.269	0.265	0.337	0.262	0.231	0.372	0.182	0.180	0.442	0.512	0.392	0.474	0.139	0.175
AdaBinsR50	134.69	81.62	0.158	0.179	0.293	0.289	0.339	0.280	0.245	0.390	0.204	0.216	0.458	0.519	0.401	0.481	0.186	0.211
AdaBinsEfficientB5⭐	100.00	85.83	0.112	0.132	0.194	0.212	0.235	0.206	0.184	0.384	0.153	0.151	0.390	0.374	0.294	0.380	0.124	0.154
DPTViT-B	83.22	95.25	0.136	0.135	0.182	0.180	0.154	0.166	0.155	0.232	0.139	0.165	0.200	0.213	0.191	0.199	0.171	0.174
SimIPUR50+no_pt	200.17	92.52	0.372	0.388	0.427	0.448	0.416	0.401	0.400	0.433	0.381	0.391	0.465	0.471	0.450	0.461	0.375	0.378
SimIPUR50+imagenet	163.06	85.01	0.244	0.269	0.370	0.376	0.377	0.337	0.324	0.422	0.306	0.289	0.445	0.463	0.414	0.449	0.247	0.272
SimIPUR50+kitti	173.78	91.64	0.312	0.326	0.373	0.406	0.360	0.333	0.335	0.386	0.316	0.333	0.432	0.442	0.422	0.443	0.314	0.322
SimIPUR50+waymo	159.46	85.73	0.243	0.269	0.348	0.398	0.380	0.327	0.313	0.405	0.256	0.287	0.439	0.461	0.416	0.455	0.246	0.265
DepthFormerSwinT_w7_1k	106.34	87.25	0.125	0.147	0.279	0.235	0.220	0.260	0.191	0.300	0.175	0.192	0.294	0.321	0.289	0.305	0.161	0.179
DepthFormerSwinT_w7_22k	63.47	94.19	0.086	0.099	0.150	0.123	0.127	0.172	0.119	0.237	0.112	0.119	0.159	0.156	0.148	0.157	0.101	0.108

Idiosyncrasy Analysis

For more detailed benchmarking results and to access the pretrained weights used in robustness evaluation, kindly refer to RESULT.md.

Create Corruption Sets

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

TODO List

• Initial release. 🚀
• Add scripts for creating common corruptions.
• Add download link of KITTI-C and NYUDepth2-C.
• Add competition data.
• Add benchmarking results.
• Add evaluation scripts on corruption sets.

Citation

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

@inproceedings{kong2023robodepth,
  title = {RoboDepth: Robust Out-of-Distribution Depth Estimation under Corruptions},
  author = {Kong, Lingdong and Xie, Shaoyuan and Hu, Hanjiang and Ng, Lai Xing and Cottereau, Benoit R. and Ooi, Wei Tsang},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2023},
}

@article{kong2023robodepth_challenge,
  title = {The RoboDepth Challenge: Methods and Advancements Towards Robust Depth Estimation},
  author = {Kong, Lingdong and Niu, Yaru and Xie, Shaoyuan and Hu, Hanjiang and Ng, Lai Xing and Cottereau, Benoit and Zhao, Ding and Zhang, Liangjun and Wang, Hesheng and Ooi, Wei Tsang and Zhu, Ruijie and Song, Ziyang and Liu, Li and Zhang, Tianzhu and Yu, Jun and Jing, Mohan and Li, Pengwei and Qi, Xiaohua and Jin, Cheng and Chen, Yingfeng and Hou, Jie and Zhang, Jie and Kan, Zhen and Lin, Qiang and Peng, Liang and Li, Minglei and Xu, Di and Yang, Changpeng and Yao, Yuanqi and Wu, Gang and Kuai, Jian and Liu, Xianming and Jiang, Junjun and Huang, Jiamian and Li, Baojun and Chen, Jiale and Zhang, Shuang and Ao, Sun and Li, Zhenyu and Chen, Runze and Luo, Haiyong and Zhao, Fang and Yu, Jingze},
  journal = {arXiv preprint arXiv:2307.15061}, 
  year = {2023},
}

@misc{kong2023robodepth_benchmark,
  title = {The RoboDepth Benchmark for Robust Out-of-Distribution Depth Estimation under Corruptions},
  author = {Kong, Lingdong and Xie, Shaoyuan and Hu, Hanjiang and Cottereau, Benoit and Ng, Lai Xing and Ooi, Wei Tsang},
  howpublished = {\url{https://github.com/ldkong1205/RoboDepth}}, 
  year = {2023},
}

License

This work is under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

Sponsor

We thank Baidu Research for the support towards the RoboDepth Challenge.

Acknowledgements

This project is supported by DesCartes, a CNRS@CREATE program on Intelligent Modeling for Decision-Making in Critical Urban Systems.