- Challenge One: Corruptions
- Challenge Two: Sensor Failures
- Rules & Conditions
- Frequently Asked Questions
The first challenge topic contains 18 real-world corruption types, ranging from the following three perspectives:
- Weather and lighting conditions, such as bright, low-light, foggy, and snowy conditions.
- Movement and acquisition failures, such as potential blurs caused by vehicle motions.
- Data processing issues, such as noises and quantizations happen due to hardware malfunctions.
| # | Task | Robust BEV Detection |
|---|---|---|
| 🚗 | Description | Evaluating the resilience of detection algorithms against diverse environmental and sensor-based corruptions |
| 📄 | Document & Instruction | track-1/README.md |
| 🏠 | Evaluation Server | https://codalab.lisn.upsaclay.fr/competitions/17135 |
| ⚙️ | Baseline Model | BEVFormer |
 |
Baseline Codebase | https://github.com/fundamentalvision/BEVFormer |
| 📊 | Baseline Results | NDS = 31.24%, mAP = 18.82% |
Hint: Participants of Track
1can refer to the above resources for more details and instructions.
| # | Task | Robust Map Segmentation |
|---|---|---|
| 🚙 | Description | Focusing on the segmentation of complex driving scene elements in BEV maps under varied driving conditions |
| 📄 | Document & Instruction | track-2/README.md |
| 🏠 | Evaluation Server | https://codalab.lisn.upsaclay.fr/competitions/17062 |
| ⚙️ | Baseline | BEVerse |
|
Baseline Codebase | https://github.com/zhangyp15/BEVerse |
| 📊 | Baseline Results | mIoU = 17.33% |
Hint: Participants of Track
2can refer to the above resources for more details and instructions.
| # | Task | Robust Occupancy Prediction |
|---|---|---|
| 🚕 | Description | Testing the accuracy of occupancy grid predictions in dynamic and unpredictable real-world driving environments |
| 📄 | Document & Instruction | track-3/README.md |
| 🏠 | Evaluation Server | https://codalab.lisn.upsaclay.fr/competitions/17063 |
| ⚙️ | Baseline | SurroundOcc |
|
Baseline Codebase | https://github.com/weiyithu/SurroundOcc |
| 📊 | Baseline Results | mIoU = 11.30% |
Hint: Participants of Track
3can refer to the above resources for more details and instructions.
| # | Task | Robust Depth Estimation |
|---|---|---|
| 🚐 | Description | Assessing the depth estimation robustness from multiple perspectives for comprehensive 3D scene perception |
| 📄 | Document & Instruction | track-4/README.md |
| 🏠 | Evaluation Server | https://codalab.lisn.upsaclay.fr/competitions/17226 |
| ⚙️ | Baseline | SurroundDepth |
|
Baseline Codebase | https://github.com/weiyithu/SurroundDepth |
| 📊 | Baseline Results | Abs Rel = 0.348, RMSE = 7.102, a1 = 0.623 |
Hint: Participants of Track
4can refer to the above resources for more details and instructions.
The second challenge topic aims to probe the 3D scene perception robustness under camera and LiDAR sensor failures:
- Loss of certain camera frames during the driving system sensing process.
- Loss of one or more camera views during the driving system sensing process.
- Loss of the roof-top LiDAR view during the driving system sensing process.
| # | Task | Robust Multi-Modal BEV Detection |
|---|---|---|
| 🚚 | Description | Tailored for evaluating the reliability of advanced driving perception systems equipped with multiple types of sensors |
| 📄 | Document & Instruction | track-5/README.md |
| 🏠 | Evaluation Server | https://codalab.lisn.upsaclay.fr/competitions/17137 |
| ⚙️ | Baseline | BEVFusion |
|
Baseline Codebase | https://github.com/mit-han-lab/bevfusion |
| 📊 | Baseline Results | NDS = 42.86%, mAP = 24.50% |
Hint: Participants of Track
5can refer to the above resources for more details and instructions.
This competition is made freely available to academic and non-academic entities for non-commercial purposes such as academic research, teaching, scientific publications, or personal experimentation. Permission is granted to use the data given that you agree:
- That the data in this competition comes “AS IS”, without express or implied warranty. Although every effort has been made to ensure accuracy, we do not accept any responsibility for errors or omissions.
- That you may not use the data in this competition or any derivative work for commercial purposes such as, for example, licensing or selling the data, or using the data with a purpose of procuring a commercial gain.
- That you include a reference to RoboDrive (including the benchmark data and the specially generated data for academic challenges) in any work that makes use of the benchmark. For research papers, please cite our preferred publications as listed on our webpage.
To ensure a fair comparison among all participants, we require:
- All participants must follow the exact same data configuration when training and evaluating their algorithms. Please do not use any public or private datasets other than those specified for model training.
- The theme of this competition is to probe the out-of-distribution robustness of autonomous driving perception models. Therefore, any use of the corruption and sensor failure types designed in this benchmark is strictly prohibited, including any atomic operation that comprises any one of the mentioned corruptions.
- To ensure the above two rules are followed, each participant is requested to submit the code with reproducible results before the final result is announced; the code is for examination purposes only and we will manually verify the training and evaluation of each participant's model.
| 🤔 | Q1: "How can I register a valid team for this competition?" |
| 🚙 | A1: To register a team, kindly fill in this Google Form. The registration period is from now till the deadline of phase one, i.e., Mar 31 '24. |
| 🤔 | Q2: "Are there any restrictions for the registration? For example, the number of team members." |
| 🚙 | A2: Each team leader should make a valid registration for his/her team. Each participant can only be registered by one team. There is no restriction on the number of team members in a team. |
| 🤔 | Q3: "Whether team members can be changed during the competition?" |
| 🚙 | A3: No. You CANNOT change the list of team members after the registration. You must register again as a new team if you need to add or remove any members of your team. |
| 🤔 | Q4: "How many tracks can I participate in?" |
| 🚙 | A4: Each team can participate in at most two tracks in this competition. |
| 🤔 | Q5: "What can I expect from this competition?" |
| 🚙 | A5: We provide the winning teams from each track with cash awards 💰 and certificates 🥇. The winning solutions will be summarized as a technical report 📖. An example of last year's technical report can be found here. |
| 🤔 | Q6: “Can I use additional data resources for model training?" |
| 🚙 | A6: No. All participants must follow the SAME data preparation procedures as listed in DATA_PREPARE.md. Additional data sources are NOT allowed in this competition. |
| 🤔 | Q7: "Can I use corruption augmentations during model training?" |
| 🚙 | A7: No. The theme of this competition is to probe the out-of-distribution robustness of autonomous driving perception models. Therefore, all participants must REFRAIN from using any corruption simulations as data augmentations during the model training, including any atomic operation that comprises any one of the corruptions in this competition. |
| 🤔 | Q8: "How should I configurate the model training? Are there any restrictions on model size, image size, loss function, optimizer, number of epochs, and so on?" |
| 🚙 | A8: We provide one baseline model for each track in GET_STARTED.md. The participants are recommended to refer to these baselines as the starting point in configuring the model training. There is no restriction on normal model training configurations, including model size, image size, loss function, optimizer, and number of epochs. |
| 🤔 | Q9: "Can I use LiDAR data for Tracks 1 to 4?" |
| 🚙 | A9: Only RAW LiDAR points data is allowed for Tracks 1 to 4 in training (e.g., generate sparse depth map). During inference, Tracks 1 to 4 are single-modality tracks that only involve the use of camera data. The goal of these tracks is to probe the robustness of perception models under camera-related corruptions. Participants who are interested in multi-modal robustness (camera + LiDAR) can refer to Track 5 in this competition. |
| 🤔 | Q10: "Is it permissible to use self-supervised model pre-training (such as MoCo and MAE)?" |
| 🚙 | A10: Yes. The use of self-supervised pre-trained models is possible. Such models may include MoCo, MoCo v2, MAE, DINO, and many others. Please make sure to acknowledge (in your code and report) if you use any pre-trained models. |
| 🤔 | Q11: "Can I use large models (such as SAM) to generate pre-training or auxiliary annotations?" |
| 🚙 | A11: No. The use of large foundation models, such as CLIP, SAM, SEEM, and any other similar models, is NOT allowed in this competition. This is to ensure a relatively fair comparing environment among different teams. Any violations of this rule will be regarded as cheating and the results will be canceled. |
| 🤔 | Q12: "Are there any restrictions on the use of pre-trained weights (such as DD3D, ImageNet, COCO, ADE20K, Object365, and so on)?" |
| 🚙 | A12: Following the most recent BEV perception works, it is possible to use pre-trained weights on DD3D, ImageNet, and COCO. The use of weights pre-trained on other datasets is NOT allowed in this competition. |
| 🤔 | Q13: "Can I combine the training and validation sets for model training?" |
| 🚙 | A13: It is strictly NOT allowed to use the validation data for model training. All participants MUST follow the nuScenes official train split during model training and REFRAIN from involving any samples from the validation set. Any violations of this rule will be regarded as cheating and the results will be canceled. |
| 🤔 | Q14: "Can I use model ensembling and test-time augmentation (TTA)?" |
| 🚙 | A14: Like many other academic competitions, it is possible to use model ensembling and test-time augmentation (TTA) to enhance the model when preparing the submissions. The participants SHOULD include necessary details for the use of model ensembling and TTA in their code and reports. |
| 🤔 | Q15: "How many times can I make submissions to the server?" |
| 🚙 | A15: For phase one (Jan. - Mar.), a team can submit up to 3 times per day and 99 times total. For phase two (Apr.), a team can submit up to 2 times per day and 49 times total. One team is affiliated with one CodaLab account only. Please REFRAIN from having multiple accounts for the same team. |
| 🤔 | Q16: " Can I use pretrained denoising or deblurring models during inference?" |
| 🚙 | A16: No. The goal of the competition is to develop a more robust perception model and using pre-trained denoising models is out of the scope of this competition. |
| 🤔 | Q17: " Can I use augmentation other than the corruption methods used in the competition?" |
| 🚙 | A17: Similar to Q7, you can use data augmentation methods that do NOT include the corruption simulation algorithms used in the competition. More details of the used corruptions can be found from this technical report. |
| 🤔 | ... |
| 🚙 | ... |
📫 Didn't find a related FAQ to your questions? Let us know (robodrive.2024@gmail.com)!
