RadarOcc: Robust 3D Occupancy Prediction with 4D Imaging Radar

3D occupancy-based perception pipeline has significantly advanced autonomous driving by capturing detailed scene descriptions and demonstrating strong generalizability across various object categories and shapes. Current methods predominantly rely on LiDAR or camera inputs for 3D occupancy prediction. These methods are susceptible to adverse weather conditions, limiting the all-weather deployment of self-driving cars. To improve perception robustness, we leverage the recent advances in automotive radars and introduce a novel approach that utilizes 4D imaging radar sensors for 3D occupancy prediction. Our method, RadarOcc, circumvents the limitations of sparse radar point clouds by directly processing the 4D radar tensor, thus preserving essential scene details. RadarOcc innovatively addresses the challenges associated with the voluminous and noisy 4D radar data by employing Doppler bins descriptors, sidelobe-aware spatial sparsification, and range-wise self-attention mechanisms. To minimize the interpolation errors associated with direct coordinate transformations, we also devise a spherical-based feature encoding followed by spherical-to-Cartesian feature aggregation. We benchmark various baseline methods based on distinct modalities on the public K-Radar dataset. The results demonstrate RadarOcc's state-of-the-art performance in radar-based 3D occupancy prediction and promising results even when compared with LiDAR- or camera-based methods. Additionally, we present qualitative evidence of the superior performance of 4D radar in adverse weather conditions and explore the impact of key pipeline components through ablation studies.

Paper

References (100)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer XRVn7/10 · confidence 3/52024-07-09

Summary

This paper introduces a novel approach for 3D occupancy prediction in autonomous driving using 4D imaging radar sensors. Traditional methods rely heavily on LiDAR or camera inputs, which are vulnerable to adverse weather conditions. RadarOcc leverages the robustness of 4D radar data, which provides comprehensive scene details even in challenging weather. The key contributions of the paper include: - **Utilization of 4D Radar Tensors (4DRT):** Unlike previous methods that use sparse radar point clouds, RadarOcc processes raw 4DRTs, preserving essential scene details and capturing comprehensive environmental data. - **Novel Data Processing Techniques:** The paper introduces Doppler bins descriptors, sidelobe-aware spatial sparsification, and range-wise self-attention mechanisms to handle the large and noisy 4DRT data. It also uses spherical-based feature encoding followed by spherical-to-Cartesian feature aggregation to minimize interpolation errors. - **Benchmarking and Evaluation:** RadarOcc is benchmarked on the K-Radar dataset against state-of-the-art methods using various modalities. The results show that RadarOcc outperforms in radar-based 3D occupancy prediction and demonstrates promising results compared to LiDAR and camera-based methods. The approach also shows superior performance in adverse weather conditions.

Strengths

1. The writing is very good, and the demo is provided for the convenience of reviewers to see the effect 2. First work use 4D imaging radar in 3D Occ. The authors propose several novel techniques, such as Doppler bins descriptors, sidelobe-aware spatial sparsification, and range-wise self-attention mechanisms, to efficiently process and utilize the voluminous and noisy 4D radar data. These methods demonstrate creativity in addressing the unique challenges posed by radar data. 3. Extensive experimental and ablation studies.

Weaknesses

1. maybe high computational demand, lack of this part description 2. Dependence on a Single Dataset 3. The experiment was limited because of the data set

Questions

1. In my understanding, the advantage of radar compared with lidar is that it can achieve good perception even in bad weather. However, because of GT procurability, author just use well-condition sequences, which i think may be unfair to lidar-based methods or vision-based methods, because the domain of radar changes little while others change a lot. It would have been better if the authors had added data on nuscenes to the baseline for comparison. Time may not be enough to complete, but this is really a point of concern to me 2. Could you provide the overlap ratio between different sensors and ROI? 3. Could you provide the overall inference speed and the speed of each module? Additionally, how much does the Doppler bins descriptor reduce the overall computational load, and do the range-wise self-attention and spherical-to-Cartesian feature aggregation introduce significant computational overhead? 4. Since Occ gt is difficult to obtain, is there any way to improve training data and generalization **I'm willing to raise the grade if address my concerns**

Rating

7

Confidence

3

Soundness

3

Presentation

4

Contribution

3

Limitations

As explained by the authors in the limitations section: 1. The model does not fully utilize the 4D radar information. 2. Due to the lack of point-wise annotations in the dataset, the semantic information is limited.

Authorsrebuttal2024-08-07

Dear Reviewer XRVn: We appreciate your detailed summary and positive comments regarding the presentation, novelty, creativity, and the experimental results of our research. Thanks a lot for providing the valuable feedback and raising insightful questions about our work. We address and answer your questions below: **Q1: Could you provide the overall inference speed and the speed of each module? Additionally, how much does the Doppler bins descriptor reduce the overall computational load, and do the range-wise self-attention and spherical-to-Cartesian feature aggregation introduce significant computational overhead?** A1: Thank you for raising this concern. Please first refer to our general answer to the **‘Computation complexity and requirement’** question in the global rebuttal. Specifically, the range-wise self-attention module is highly efficient, with a runtime of only 2.5ms. The spherical-to-Cartesian feature aggregation initially costs 72ms per frame. However, this can be optimized to 29.7ms by reducing the number of layers. This module not only converts the spherical feature volume to a Cartesian one without interpolation errors but also enhances the encoding of voxel feature relationships. The Doppler bins descriptor significantly reduces the data volume of raw 4D radar tensors (4DRTs) by a factor of D/8, where D equals 64 in our case. This reduction decreases the input data volume to 1/8 of its original size. This reduction not only shortens loading times (from disk/sensor to processors) but also lessens the overall computational burden during spherical-based feature encoding. **Q2: Dependence on a Single Dataset. The experiment was limited because of the data set** A2: Thanks for highlighting this concern. Please refer to our general answer to the **‘Dependence on a single dataset’** answer in the global rebuttal. **Q3: Author just use well-condition sequences, which I think may be unfair to lidar-based methods or vision-based method** A3: Thanks for this thoughtful feedback. We are not sure if we understand this question exactly. If you were concerned that the training data used to train baselines method does not include sequences collected from adverse weathers, which causes worse performance due to the domain change of LiDAR and camera data, we are happy to answer this question as follows. Currently, it is hard to resolve this problem as the LiDAR data is inevitably affected by the adverse weather. For example, water droplets or snowflakes can scatter or absorb LiDAR beams, reducing the effective range of LiDAR and inducing noise in the data. As a result, the occupancy labels obtained from accumulated LiDAR data under adverse weather have low fidelity. The nuScenes dataset also faces a similar issue where occupancy labels are affected by adverse weather, despite having annotated bounding boxes. Adding new nuScenes data collected under adverse weather to train our LiDAR and vision-based baselines might not be feasible. This approach could potentially hinder network training because of noisy labels rather than improve adaptation to adverse weather. Moreover, as nuScenes does not provide 4D radar tensor data, it’s unfair for our radar-based methods because they cannot benefit from the new data samples. This situation highlights a significant research question: how can we generate reliable 3D occupancy labels under adverse weather conditions when LiDAR measurements are noisy? We acknowledge the importance of this question and plan to address it in our future work. **Q4: Could you provide the overlap ratio between different sensors and ROI?** A4: Thanks for this comment. Our evaluation focuses on the overlap area between the horizontal field of view (FoV) of all sensors and our defined RoI to minimize potential data discrepancies beyond the FoV. Specifically, the overleap hFoV of K-Radar sensor suite is 107◦, symmetrically distributed around the front axis. The ratio of between the final evaluation area and our RoI can be calculated as: $$ 1 - \frac{\cot\left(\frac{107^\circ}{2}\right)}{4} \approx 0.812 $$ which means 81.2% area of our RoI is taken into account for our final evaluation area. **Q5: Any way to improve training data and generalization** A5: Thanks for this insightful question. It aligns with our research interest in efficiently training 3D occupancy prediction models. Some recent works are exploring self-supervised methods in occupancy prediction. One notable example is EmerNeRF (https://emernerf.github.io/), which can represent highly dynamic scenes in a self-sufficient manner. This approach has the potential to label occupancy ground truth using RGB cameras without requiring human annotation efforts.

Reviewer XRVn2024-08-13

Thank you to the authors for their response. I have carefully reviewed the feedback from the other reviewers as well as the corresponding rebuttal. While the rebuttal has addressed most of my concerns, the limited experimental results prevent me from raising my score. I believe the score I have given is appropriate for this submission. It is a commendable piece of work, and I look forward to seeing more research in this area.

Reviewer f9yn4/10 · confidence 4/52024-07-10

Summary

This paper leverages recent advancements in automotive radar technology and introduces a novel approach that utilizes 4D imaging radar sensors for 3D occupancy prediction. The proposed method incorporates Doppler bin descriptors, sidelobe-aware spatial sparsification, and range-wise self-attention mechanisms to effectively manage and mitigate the noise present in 4D radar data.

Strengths

This paper presents a compelling algorithm designed to enhance radar-based 3D occupancy prediction by incorporating Doppler information and sidelobe-aware techniques. The authors propose an innovative pipeline that effectively reduces data volume, mitigates sidelobe measurements, and utilizes interpolation-free feature encoding and aggregation.

Weaknesses

Please consult the question section for further information.

Questions

1. Dataset Limitation: The experimental analysis is thorough, but it is conducted solely on the K-Radar dataset. There is a concern that the method may be overfitted to this specific dataset. Could the authors provide additional examples using other datasets to demonstrate the generalizability of the method? 2. Computational Complexity: The methodology includes both an attention mechanism and multi-scale 3D convolutional blocks. A potential issue is whether this structure significantly increases the computational requirements. Could the authors provide a comparison of model size and computational time to address this concern? 3. Selection Criteria: The method selects the Top-3 power values and their corresponding indices. How does this choice affect performance? Would increasing or decreasing the number of selected values significantly impact the results? 4. Failure Cases: Are there any documented failure cases of the method? Considering that Doppler effects are most pronounced when the relative velocity is aligned with the wave direction, how does the method perform in scenarios where this condition is not met? 5. Would the code be published together with the paper?

Rating

4

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Please consult the question section for further information.

Authorsrebuttal2024-08-14

Rebuttal follow up

Dear reviewer f9yn, As the discussion period is approaching its end, we kindly invite you to review our detailed rebuttal. We believe it addresses the concerns you raised in your review. If you find our work satisfactory, we would greatly appreciate it if you could consider raising your score to a positive one. We also welcome any further comments you may have. Thank you again for your time in reviewing our paper.

Reviewer 7z3i5/10 · confidence 3/52024-07-12

Summary

This paper introduces a 3D occupancy prediction method that, unlike previous radar-based approaches, utilizes 4D imaging radar to leverage additional information. To harness the potential of this under-explored 4D data, the paper tackles challenges such as the large size of raw 4D radar data, inherent noise, and discrepancies in coordinate systems. The proposed method results in improved reconstruction accuracy compared with other baselines.

Strengths

The motivation to further exploit the 4D radar data for occupancy estimation is logical. This paper represents pioneering work in addressing this issue. The paper is well-crafted, featuring clear figures and well-organized content. The performance under adverse weather conditions is noteworthy, underscoring a critical perception task for the safety of current autonomous vehicles (AVs).

Weaknesses

Comparing the proposed method to only one published baseline and relying on a single dataset raises concerns about the persuasiveness of the findings. As a learning-based model, the paper misses an evaluation of the generalization of the learning formulation to datasets out of the domain. This oversight is crucial for assessing the model's applicability across various real-world scenarios.

Questions

Can the proposed learning formulation be extended to incorporate radar methods with other modalities, such as images and lidar? What is the trade-off relationship between compressing the input data (originally 500MB) and its impact on performance? Providing this information could assist readers in balancing efficiency and accuracy. MISC Line 110: The line distance is unusual.

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The experiments are conducted solely on the K-Radar dataset, which is the only autonomous driving dataset containing 4D radar data. This situation highlights a significant issue: the scarcity of 4D data compared to other perception methods (e.g., images, normal radar data), which limits its applicability in broader contexts.

Reviewer v56z5/10 · confidence 5/52024-07-13

Summary

It is proposed to utilize 4D imaging 8 radar sensors for 3D occupancy prediction by directly processing the 4D radar tensor, thus preserving essential scene details. RadarOcc innovatively addresses the challenges associated with the voluminous and noisy 4D radar data by employing Doppler bins descriptors, sidelobe-aware spatial sparsification, and range-wise self-attention mechanisms. The demonstration of the RadarOcc’s state-of-the-art performance in radar-based 3D occupancy prediction was carried out on K-Radar dataset.

Strengths

It is well-accepted that radar raw data could provide more information for perception tasks in autonomous driving. This submission follows the same direction to utilize 4D imaging radar measurement.

Weaknesses

It is not the first paper to utilize the 4D imaging radar raw measurement for perception. As mentioned by the authors, they focus on semantics, rather than typical road users detection and classification. For the reference to 4D imaging radar, please consider cite the following paper S. Sun and Y. D. Zhang, "4D Automotive Radar Sensing for Autonomous Vehicles: A Sparsity-Oriented Approach," in IEEE Journal of Selected Topics in Signal Processing, vol. 15, no. 4, pp. 879-891, June 2021. For general automotive radar contribution to autonomous driving, please consider cite the following paper S. Sun, A. P. Petropulu and H. V. Poor, "MIMO Radar for Advanced Driver-Assistance Systems and Autonomous Driving: Advantages and Challenges," in IEEE Signal Processing Magazine, vol. 37, no. 4, pp. 98-117, July 2020.

Questions

It seems to be a trade-off in the sidelobe-aware spatial sparsification process. Is it interesting to know whether important feature information is lost due to this sparsification. For example, targets with weak reflections, such as the information of road curbs might be lost.

Rating

5

Confidence

5

Soundness

3

Presentation

2

Contribution

2

Limitations

As mentioned before, the proposed work does not apply to the general road users detection and classification.

Authorsrebuttal2024-08-14

Dear reviewer v56z, As the discussion period is approaching its end, we kindly invite you to review our detailed rebuttal. We believe it addresses the concerns you raised in your review. We also welcome any further comments you may have. Thank you again for your time in reviewing our paper.

Reviewer kdzm5/10 · confidence 4/52024-07-13

Summary

The paper presents RadarOcc, a method that enhances 3D occupancy prediction for autonomous vehicles using 4D imaging radar. It directly processes the 4D radar tensor, aiming to overcome the limitations of sparsity and noise associated with conventional radar processing. The methodology introduces techniques such as Doppler bins descriptors and sidelobe-aware spatial sparsification to improve data integrity and scene detail retention. The evaluation is conducted on the K-Radar dataset, with RadarOcc demonstrating superior performance compared to traditional LiDAR and camera-based methods, particularly under adverse weather conditions.

Strengths

The utilization of 4D imaging radar data directly (instead of converting it to sparse point clouds) allows for a more comprehensive environmental perception. It is also good to consider different weather conditions.

Weaknesses

1. The processing of dense radar tensors can be computationally expensive, potentially limiting real-time application in less powerful onboard systems. 2. The current framework processes single-frame data, which may not fully capture the dynamic nature of driving environments, possibly affecting prediction reliability in highly dynamic scenarios. 3. As this study focuses on solving real problems in autonomous driving, it is important to thoroughly justify why 4D imaging radar has advantages over LiDAR in addressing issues like occupancy prediction. Can methods related to LiDAR really not solve these problems in various environments? In fact, many autonomous vehicles not only refrain from using mm-wave radar but also do not use LiDAR, managing to solve most issues using only vision.

Questions

1. How does the RadarOcc system perform in terms of real-time processing, and what are the computational requirements? 2. Could the method be integrated with temporal data to predict dynamic changes in the environment, and if so, what modifications would be necessary? 3. Are there specific scenarios or environments where RadarOcc's performance might be significantly reduced, such as urban canyons or areas with high radar interference? 4. How does the system handle objects with low radar cross-section, such as pedestrians or animals?

Rating

5

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

1. The method's high computational demand may not scale well to lower-end processors commonly used in commercial vehicles. 2. The lack of temporal modeling could limit the system's predictive capabilities in highly dynamic environments. 3. It remains unclear how well the method generalizes across different radar systems or configurations not represented in the K-Radar dataset.

Authorsrebuttal2024-08-14

Rebuttal follow up

Dear reviewer kdzm, As the discussion period is approaching its end, we kindly invite you to review our detailed rebuttal. We believe it addresses the concerns you raised in your review. We also welcome any further comments you may have. Thank you again for your time in reviewing our paper.

Authorsrebuttal2024-08-13

Rebuttal Response Follow-Up

Dear our reviewers, I am writing to follow up on the rebuttal we submitted for our paper. We understand the review process can be time-consuming but as the discussion deadline is approaching, we would appreciate any updates or feedback at your earliest convenience so that we could do some last-minute clarification if needed. We have not yet received any feedback and want to ensure that our reviewers have had the opportunity to read our rebuttal, does our rebuttal answer all your questions? Best regard, Submission391 authors

Reviewer 7z3i2024-08-13

Comment

Thanks for the authors' response to my concerns. I have no further questions and feel I can keep my positive ratings for the paper.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC