RETR: Multi-View Radar Detection Transformer for Indoor Perception

Indoor radar perception has seen rising interest due to affordable costs driven by emerging automotive imaging radar developments and the benefits of reduced privacy concerns and reliability under hazardous conditions (e.g., fire and smoke). However, existing radar perception pipelines fail to account for distinctive characteristics of the multi-view radar setting. In this paper, we propose Radar dEtection TRansformer (RETR), an extension of the popular DETR architecture, tailored for multi-view radar perception. RETR inherits the advantages of DETR, eliminating the need for hand-crafted components for object detection and segmentation in the image plane. More importantly, RETR incorporates carefully designed modifications such as 1) depth-prioritized feature similarity via a tunable positional encoding (TPE); 2) a tri-plane loss from both radar and camera coordinates; and 3) a learnable radar-to-camera transformation via reparameterization, to account for the unique multi-view radar setting. Evaluated on two indoor radar perception datasets, our approach outperforms existing state-of-the-art methods by a margin of 15.38+ AP for object detection and 11.91+ IoU for instance segmentation, respectively. Our implementation is available at https://github.com/merlresearch/radar-detection-transformer.

Paper

Similar papers

Peer review

Reviewer XrXa6/10 · confidence 4/52024-06-15

Summary

This work introduces a novel method RETR for indoor object detection and segmentation based on multi-view radar heatmaps. RETR extends the popular DETR framework and incorporates modifications specific to multi-view radar perception, such as depth-prioritized feature similarity via TPE, a trip-plane loss, and a learnable radar-to-camera transformation. Experiments on two datasets demonstrate the effectiveness of RETR for object detection and segmentation.

Strengths

1. This work utilizes multi-view radar heatmaps as input for indoor human perception, which has broad application for privacy-aware indoor sensing and monitoring. 2. This work extends image-based DETR to multi-view radar-based RETR, establishing a new baseline for radar-based human detection and segmentation. The proposal of tunable positional embedding is interesting. It can enable the adjustment of different axes for positional embeddings. 3. The experiments are comprehensive with two tasks validated on two tasks. The details of hyperparameters and implementation details are adequate, making this paper a good reference for future works.

Weaknesses

1. Most commercial radars have 2D virtual antenna arrays, e.g., 16*8, instead of only a pair of 1D antenna arrays. Is there any specific reason for considering such kinds of radar antenna arrays? Normally, the 2D heatmaps (range-elevation, range-azimuth) are generated via projecting the 3D radar cube data in two views. The explanation in Section 2 does not fit the real process of multi-view radar heatmap generation. If this work considers perception with such a unique radar antenna array, better to introduce what types of radar the dataset used. 2. In the experiments, only radar heatmap-based methods are used for comparison. Better to incorporate the results of radar point cloud-based and camera-based approaches for a more complete comparison. The survey of related work is also not complete, ignoring recent works in radar-based object detection for autonomous driving [1-3], which also use radar heatmaps: [1] Liu, Yang, et al. "Echoes beyond points: Unleashing the power of raw radar data in multi-modality fusion." Advances in Neural Information Processing Systems 36 (2024). [2] Paek, Dong-Hee, Seung-Hyun Kong, and Kevin Tirta Wijaya. "K-radar: 4d radar object detection for autonomous driving in various weather conditions." Advances in Neural Information Processing Systems 35 (2022): 3819-3829. [3] Skog, Mikael, et al. "Human Detection from 4D Radar Data in Low-Visibility Field Conditions." arXiv preprint arXiv:2404.05307 (2024).

Questions

1. At line 28, it is better to explain why radar heatmaps are more preferred for challenging perception tasks. At line 31, it's not necessary to mention if one work is publicly accessible for the introduction section. 2. The relationship and different between previous works in radar-based detection and segmentation, e.g., RFMask and RETR is not well explained. 3. Better to show the tokenization step in Figure 3 to aid in understanding the pipeline. 4. In line 135, equation 3 should be Equation 3. 5. How did the author implement the top-K feature selection step? And how to supervise such a selection step to ensure the selected features are significant for the detection task? 6. In line 229, what's the reason why the author claimed that the calibration may be only accurate for a limited interval of depth and angles. 7. The MMVR dataset can not be found online. 8. Please provide captions for table 1 and 2 to increase the readability.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

NA.

Reviewer kPes6/10 · confidence 2/52024-07-12

Summary

The primary content of this paper is an introduction to a multi-view radar detection transformer algorithm (RETR) for indoor perception. The algorithm achieves effective object detection in indoor environments by utilizing multi-view radar data and combining self-attention mechanisms and cross-attention mechanisms. The author validates the performance improvement of the algorithm through experiments and discusses its application in indoor perception.

Strengths

1.The idea is helpful for the field of view indoor multi-view radar perception. 2.The paper is clear and easy to follow. 3.Rigorous ablation studies were conducted, providing evidence of the proposed method's efficacy. 4.RETR, based on the original DETR, has achieved significant performance improvements on indoor multi-view radar datasets with heatmap input.

Weaknesses

1.Radar perception datasets primarily utilize point clouds as the data form. Methods based on heatmaps for indoor multi-view radar perception are not yet sufficient. The author mainly compares with customized DETR and RFMask, but the comparison is insufficient. Have the authors considered conducting further comparison experiments on the HUPR [1] dataset? If the authors can further demonstrate the method's generalization ability, I would consider giving a higher score. 2.The method presented is an improvement on the DETR and combines existing methods in a novel way. While innovative, it's not highly original for NeurIPS. 3.Research on indoor multi-view radar perception using heatmaps is relatively scarce. The author is unable to provide source code, potentially limiting contributions to this field. [1] Lee, Shih-Po, et al. "Hupr: A benchmark for human pose estimation using millimeter wave radar." Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2023.

Questions

1.Could you provide a detailed description of the design of the Top-k selector? 2.Can you provide a more detailed analysis of the impact of the proposed RETR on real-time performance?

Rating

6

Confidence

2

Soundness

3

Presentation

3

Contribution

3

Limitations

NA

Reviewer H5qV7/10 · confidence 5/52024-07-13

Summary

In this paper, the authors propose Radar dEtection TRansformer (RETR), an extension of the popular DETR architecture, tailored for multi-view radar perception. RETR inherits the advantages of DETR, eliminating the need for hand-crafted components for object detection and segmentation in the image plane.

Strengths

Two radars with different directions are deployed. Transformer is applied for segmentation.

Weaknesses

For indoor perception with radar, multi-path is expected. It would be good to add a paragraph to discuss this issue and how to mitigate multi-path or its impact to the perception/segmentation performance. The motivation of applying transformer for indoor segmentation is not well discussed. What are the benefits and unique challenges of applying transformer to radar based indoor segmentation? For the introduction of Generation of Radar Heatmaps, 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

Is there any interference (cross-path signals) between the two radars deployed in vertical and horizontal directions?

Rating

7

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

Usually a large amount of high quality radar data is required to train the transformer. It is highly recommended to carry out validation with different amount of training data.

Reviewer DJHb6/10 · confidence 4/52024-07-15

Summary

The paper introduces Multi-View Radar Detection Transformer for indoor object detection. Inspired from DETR, the authors propose and end-to-end RETR to detect objects from the radar inputs. To improve the feature association across the two radar views, the authors introduce a new Tunable Positional Embedding. The proposed approach achieves the solid performance on the standard benchmarks of radar-based indoor object detection.

Strengths

- The paper is well-written and easy to follow. - The proposed approach is simple yet well-motivated. The proposed RETR eliminates the cumbersome design of the detection network for radar inputs. - The analysis of Tunable Positional Embedding Section 4.3 is comprehensive and well-motivated. - The proposed method achieves strong experimental results on MMVR and HIBER datasets.

Weaknesses

- I am wondering how the Top-K Feature selection impacts the performance of the network. How can we choose the number of K? It seems that there is no experiment to validate the choice of K. It will be better if the authors conduct an ablation study to explore the effectiveness of choosing K. - What is the computation cost of the proposed RETR model? It will be better if the authors report the inference time/computational cost of the proposed model - For the learnable radar-to-camera coordinate transformation, while I acknowledge the performance improvement according to this proposed model, I have several questions related to this module. - - Does the entire dataset use the same set of learnable vector $\omega$ and translation vector $t$? Or each radar sample in the data will have a different vector $\omega$ and translation vector $t$? - - How can we verify that the learnable radar-to-camera coordinate transformation is accurate? Is there any way to evaluate it using the ground truths?

Questions

Please refer to my weakness section.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have discussed the limitations and broader impact in the paper.

Reviewer kPes2024-08-12

Thank you for your response and the detailed clarification of my comments. The explanation regarding the use of the HuPR dataset and the challenges faced in obtaining labels such as Bounding Boxes is indeed insightful. This complexity could potentially limit the scope of your current evaluation, especially during the rebuttal period. Your explanation of the Top-K selection mechanism and the related experiments on inference time further strengthen the argument for the effectiveness of RETR. Overall, I believe your research contributions are significant and closely related to the field. The decision to open-source the code will have a positive impact on the research community for radar perception tasks. Thank you for your efforts in addressing my concerns, and I have decided to adjust my recommendation to a score of 6.

Authorsrebuttal2024-08-12

We sincerely appreciate the reviewer's time and effort in reviewing our rebuttal, and we're delighted to see that it has been positively received. We will certainly consider the suggested improvements for the updated paper.

Reviewer XrXa2024-08-12

Thank the authors for providing such detailed response to my comments. I am looking forward to see more explanantion regarding the radar configuration and generation of the radar heatmaps in the revised version. Regarding perception methods based on radar heatmap, we find two more recent works: [1] Kong, Seung-Hyun, Dong-Hee Paek, and Sangyeong Lee. "RTNH+: Enhanced 4D Radar Object Detection Network using Two-Level Preprocessing and Vertical Encoding." IEEE Transactions on Intelligent Vehicles (2024). [2] Ding, Fangqiang, Xiangyu Wen, Yunzhou Zhu, Yiming Li, and Chris Xiaoxuan Lu. "RadarOcc: Robust 3D Occupancy Prediction with 4D Imaging Radar." arXiv preprint arXiv:2405.14014 (2024). Hope the citation of them could improve the rigorousness of your related work. Overall, thank you for your efforts in addressing my concerns, and I have decided to adjust my recommendation to a score of 6.

Authorsrebuttal2024-08-12

We sincerely appreciate the reviewer's time and effort in reviewing our rebuttal, and we're delighted to see that it has been positively received. We will certainly consider the suggested references to update our related work. Additionally, we kindly encourage the reviewer to update the score to reflect the stated intentions.

Reviewer DJHb2024-08-12

Feedback to Author Rebuttal

Thank the authors for the good rebuttal. It has addressed my concerns. I hope that you can include these answers and experiments in your revised version. Therefore, I decided to increase my score to 6.

Authorsrebuttal2024-08-12

We sincerely appreciate the reviewer's time and effort in reviewing our rebuttal, and we're delighted to see that it has been positively received. We will certainly consider the suggested improvements for the updated paper.

Authorsrebuttal2024-08-13

We appreciate the raised ethical concerns and would like to first address them in the context of our work on RETR. We will then respond to your concerns regarding privacy and security related to HIBER and MMVR to the best of our ability, as these are separate publications. **RETR** - In training RETR with the MMVR and HIBER datasets, we only utilized radar heatmaps (without any RGB images) and annotation labels such as bounding boxes (BBox) and segmentation masks. Fig.11 provides an illustration of BBoxes and masks in HIBER. From the heatmaps and labels, it is difficult to directly identify individuals or infer personal attributes like gender and weight. While the mask and BBox might indicate height, the data used is limited in revealing sensitive information. We included de-faced RGB images from MMVR only for illustration purposes by requesting selected frames from the MMVR authors. We did not have any RGB images from the HIBER dataset for illustration. - Although this RETR paper did not involve creating a new radar perception dataset, we emphasized in the "Broader Impacts" section at the end of the paper that while indoor radar perception models may open up applications for good uses such as elderly care and fall detection, it is crucial that the resulting perception results are kept secure and private, ensuring they are not misused to infer attributes like gender, size, or height without permission. Alongside the subsequent discussions on the MMVR and HIBER datasets, we plan to expand the "Broader Impacts" section to further raise awareness of the potential harms and misuse of radar perception models. **Questions regarding HIBER and MMVR** - *How was the data de-identified...?* - The HIBER dataset and the MMVR dataset do NOT contain RGB images. Only radar heatmaps and annotation labels such as BBoxes, keypoints, and segmentation masks, are included. Due to the nature of radar heatmaps and these annotations, it is difficult to directly identify individuals, ensuring a level of de-identification in the data. - *What measures...?* - Although we cannot speak on behalf of the authors of these datasets, based on the information provided in their papers, we have learned the following: - **No RGB Images Released:** No RGB images are included in the publicly available versions of the datasets. A small number of RGB images were used in the HIBER and MMVR papers, but in the MMVR paper, these images were all de-faced to protect participant identities. - **Informed Consent and Data De-identification:** The MMVR paper includes a statement in Sec.6 that outlines their ethical procedures. Participants were informed at the beginning of each data session about the experiment's goals, procedures, potential exposure to high-frequency radio waves, and the use of cameras. They were also notified that de-identified data would be made publicly accessible for research purposes, and all participants signed a consent form. - We did not find specific statements regarding data collection protocols in the HIBER paper, but it is noteworthy that on the MMVR dataset webpage, participant faces are masked. - *Do the datasets remove...?* - The major measure is to NOT include RGB images in both datasets. This significantly reduces the risk of visual identification. Also, it appears to us that the subjects are anonymized in the dataset. - *Are there any methods or...?* - There are safeguards in place for both the HIBER and MMVR datasets, according to their papers and dataset websites: - **HIBER**: Any user of the HIBER dataset is required to sign an agreement that includes clauses limiting the use of the dataset only for scientific purposes and pointing out potential misuse and harm of using their dataset. This obligation serves as a deterrent against inappropriate uses, including unauthorized indoor surveillance. - **MMVR**: The authors of the MMVR paper have explicitly addressed the ethical implications of using radar perception technologies for purposes such as classifying and estimating personal attributes. In Sec.7 of their paper, they raise awareness of the potential misuse of the dataset, noting: "Although radar perception has fewer privacy concerns than cameras, MMVR can potentially be utilized to classify and estimate attributes of subjects such as gender, size, height, and gait. It may also be used to advance technologies for indoor surveillance without acknowledgment or permission." This acknowledgment helps to inform users of the ethical considerations and potential risks associated with the dataset, encouraging responsible use. Since our paper focuses on training a radar perception model, not the dataset itself, we’re considering legal clauses similar to or stronger than those in the HIBER dataset for users who are interested in using our code and model weights. We’ll also expand the "Broader Impacts" section to highlight potential harms and misuse of these models.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC