4DR P2T: 4D Radar Tensor Synthesis with Point Clouds

Dong-Hee Paek; Seung-Hyun Kong; Woo-Jin Jung

REVIEW 3 major objections 4 minor 25 references

4DR P2T: 4D Radar Tensor Synthesis with Point Clouds

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that a conditional generative adversarial network can regenerate dense 4D radar tensors from sparse point clouds, reaching an average PSNR of 30.39 dB and SSIM of 0.96, and that percentile-1% point clouds offer the best…

desk verdict A plausible new task—radar tensor reconstruction from point clouds without LiDAR—but the paper's headline conclusion about which point cloud method is best for deep learning is not demonstrated by the evidence. read the letter →

arxiv 2502.05550 v1 pith:RO673GHC submitted 2025-02-08 cs.CV

Woo-Jin Jung , Dong-Hee Paek , Seung-Hyun Kong This is my paper · ORCID

classification cs.CV

keywords 4DradarpointcloudtensorgenerationconditionalGANsparseconvolutionautonomousdrivingK-Radardatacompression

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper aims to show that a conditional generative adversarial network (cGAN) can regenerate dense radar tensors from sparse 4D radar point clouds, turning the point cloud into an encoded form and the model into a decoder. On the K-Radar dataset, the proposed 4DR P2T model reaches an average peak signal-to-noise ratio of 30.39 dB and a structural similarity index of 0.96 across several point-cloud generation methods. The paper also claims that percentile-5% point clouds give the best tensor reconstruction, while percentile-1% point clouds give the best trade-off between data-volume reduction and reconstruction performance for deep learning. If these claims hold, radar perception pipelines could store or transmit point clouds and regenerate dense tensors on demand, and point-cloud extraction methods could be ranked by reconstruction fidelity.

What carries the argument

The load-bearing machinery is an encoder-decoder generator with skip connections: the encoder processes sparse point clouds with 3D sparse convolutions at downsampling stages and submanifold sparse convolutions at full-resolution stages, while the decoder uses 3D dense convolutions to produce a complete dense tensor. A multi-scale 3D discriminator judges the output, and the training objective combines a conditional adversarial loss, an L1 loss, and a perceptual loss. The model treats the point cloud as an encoded version of the tensor, so its reconstruction quality serves as a measuring stick for how much environmental information each point-cloud generation method preserves.

What would settle it

Train a downstream object detector on tensors regenerated from percentile-1% point clouds and compare its performance against the same detector trained on raw tensors; if detection performance does not follow the PSNR/SSIM ranking, the proxy assumption fails. Alternatively, compute PSNR and SSIM directly on the full 3D tensors without height-pooling and check whether percentile-5% still ranks first.

Watch

Extended reading notes

Core claim

The central discovery is that a cGAN composed of a 3D sparse-convolution encoder and a 3D dense-convolution decoder can reconstruct dense 4D radar tensors from sparse point clouds well enough to rank different point-cloud generation methods. Evaluated on K-Radar, the model achieves its best reconstruction with percentile-5% point clouds (34.43 dB PSNR, 0.98 SSIM), and the percentile-1% method attains the highest deep-learning efficiency score of 0.48, indicating the most favorable ratio of reconstruction fidelity to point-cloud density. The paper interprets tensor-generation performance as a proxy for how well a point cloud preserves environmental information, which is what makes the comparison of extraction methods meaningful.

Load-bearing premise

The paper's load-bearing premise is that PSNR and SSIM computed after mean-pooling the generated 3D tensor along the height axis correctly indicate how much environmental information a point cloud preserved and how suitable the reconstructed tensor is for downstream deep learning; if this proxy is wrong, the recommendation of the 1% percentile method loses its basis.

Editorial extensions

If this is right

The 4DR P2T model can act as a decoder that regenerates dense radar tensors from sparse point clouds, enabling point-cloud storage or transmission with later tensor reconstruction.
Percentile-5% point clouds should be preferred when reconstruction quality is the priority, since they yield the highest PSNR and SSIM.
Percentile-1% point clouds offer the best performance-to-data-volume ratio, making them a strong choice for training deep learning models under bandwidth or memory limits.
The model's current form excludes Doppler information to keep computational costs low, so extending it to full four-dimensional tensors is a natural next step.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If reconstruction quality is a valid proxy for information preservation, the same point-to-tensor training could be applied to radar datasets that only provide point clouds, effectively manufacturing tensor supervision.
The method suggests a practical radar data-compression pipeline: retain the 1% percentile point cloud and regenerate the dense tensor on demand, trading some fidelity for lower transmission cost.
The paper does not validate the proxy assumption with a downstream object-detection experiment on the reconstructed tensors; that experiment would directly test whether the PSNR/SSIM ranking matters for deep learning.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

Here's the quick take: this paper proposes a genuinely new task—reconstructing dense 4D radar tensors from the same sensor's point clouds, with no LiDAR supervision—and builds a sensible cGAN architecture to do it. The K-Radar dataset makes this possible, and the authors deserve credit for noticing the gap and for describing their model and training clearly enough to reimplement.

The empirical comparison across CFAR and percentile point-cloud densities is useful as a first exploration. But the central claim—that the 1% percentile method is "well-suited for deep learning training"—rests on much weaker evidence. There are no baselines at all: no simple voxelization of the input point cloud, no deterministic regression, no UNet without the GAN. Without those, we can't tell whether the learned model adds anything over a cheap interpolation. There are also no error bars or significance tests, so differences like 31.66 vs. 30.00 dB could be noise.

The DES metric is the biggest soft spot. It divides normalized PSNR and SSIM by point cloud density, which by construction rewards sparser point clouds. So the finding that 1% beats 5% is baked into the metric's definition, not demonstrated by task performance. The other evaluation choice—mean-pooling the 3D tensor along height before computing PSNR/SSIM—can hide elevation errors and is dominated by empty regions. And the paper's stated motivation, that reconstructed tensors are "suitable for deep learning," is never tested by a downstream detector or segmentation network.

The abstract's average PSNR of 30.39 dB is also a bit cherry-picked: it averages over very different density settings, and comparing that number to wireless communication thresholds is not meaningful evidence.

All that said, the paper is not fundamentally broken. The task is real, the architecture is reasonable, and the authors are transparent about their choices. Self-citation is heavy but not inappropriate given they created the dataset.

Who is this for? Researchers working on radar representation learning or sensor compression might find it a useful starting point, but only as a benchmark to beat. It deserves serious peer review, because the core idea is worth testing properly—but it would need major revisions: baselines, uncertainty estimates, downstream validation, and a defensible metric. As it stands, I would not cite it yet.

Recommendation: send to review, but require the missing comparisons before publication.

Referee Report

3 major / 4 minor

Summary. This paper proposes 4DR P2T, a conditional generative adversarial network (cGAN) that synthesizes dense 3D radar tensors from sparse 4D radar point clouds. The model uses 3D sparse convolutions in the encoder and 3D dense convolutions in the decoder, with a multi-scale discriminator, and is trained on the K-Radar dataset with ground-truth tensors as supervision. The authors evaluate the method using PSNR and SSIM computed after mean-pooling the output along the height axis, and introduce a deep-learning efficiency score (DES) that divides normalized quality scores by point cloud density. The paper reports an average PSNR of 30.39 dB and SSIM of 0.96, and concludes that the 5% percentile point cloud generation method gives the best tensor quality while the 1% percentile method is best suited for deep learning applications.

Significance. If the claims were fully substantiated, the paper would offer a useful data-format conversion tool for 4D radar perception and an empirical comparison of point cloud generation methods. The work has several strengths: it uses the public K-Radar dataset, evaluates against held-out ground-truth tensors, and clearly describes the architecture. The reported reconstruction scores are computed against held-out ground truth. However, the central conclusion that the 1% percentile is optimal for deep learning rests on an unvalidated, density-normalized metric, and the evaluation lacks baseline comparisons and 3D-aware quality measures.

major comments (3)

[Section IV.C, Eq. (6)] The DES metric divides the normalized PSNR and SSIM by D(i), the point cloud density (in percent). Because D(i) is in the denominator, the metric systematically rewards sparser point clouds; for example, the 1% percentile row in Table I achieves the highest DES (0.48) mainly because its PCD (1.11) is much smaller than the 5% row's PCD (4.46), despite having lower raw PSNR/SSIM than the 5% row. The conclusion in Section IV.D that the 1% method is 'well-suited for deep learning model training' is therefore an artifact of the metric's construction unless DES is validated against an actual downstream task. The paper should either demonstrate such validation (e.g., object detection or segmentation performance on the reconstructed tensor) or soften the claim.
[Section IV.C] The evaluation converts the 3D generated tensor to a 2D image by mean pooling along the height axis before computing PSNR and SSIM. This can mask errors in the elevation dimension and is dominated by large empty regions of the BEV projection, so two tensors with very different 3D fidelity can receive similar scores. The paper's central proxy claim, stated in Section I ('tensor generation performance ... serves as a proxy for assessing how well a given point cloud preserves environmental information'), is therefore not established. The authors should additionally report 3D voxel-wise metrics or demonstrate that height-pooled quality correlates with downstream perception performance.
[Section IV.B and Table I] The evaluation lacks any baseline comparison. The paper does not report how simple alternatives—such as direct voxelization of the input point cloud, interpolation, a deterministic UNet without the adversarial loss, or a non-learned fill—perform on the same metrics. Without such baselines, the abstract's claim that results 'validate the effectiveness of the 4DR P2T model' is unsupported, as the reported average PSNR of 30.39 dB has no reference point. Adding at least a voxelization baseline and an ablation that removes the GAN objective would clarify the model's contribution.

minor comments (4)

[Section IV.D] The 'average PSNR across all methods is 30.39dB' is the mean of the six method-level PSNR values in Table I, not an average over test frames. The paper should clarify this and report per-frame statistics with standard deviations to substantiate the claimed stability.
[Section IV.C] The text defines D(i) as a 'ratio' of detected points to total possible points, but the table reports PCD as percentages (e.g., 1.11, 4.46). The inconsistency in units should be fixed, since the absolute DES values depend on the scale of D(i).
[Section III.C] The perceptual loss is not fully specified; the paper should name the pre-trained network (e.g., VGG) and the layers used for feature extraction to enable reproducibility.
[Section IV.D] The use of a 20–25 dB 'acceptable quality' threshold from wireless communication studies [24], [25] is not directly applicable to radar tensor reconstruction, and the paper should provide a more relevant justification for interpreting the reported PSNR as 'sufficiently high performance.'

Circularity Check

1 steps flagged · score 4.0 of 10

Core cGAN reconstruction is independent, but the '1% percentile optimal for deep learning' conclusion is a restatement of the self-defined DES metric rather than an empirical result.

self definitional [Section IV.C (Metrics, Eq. 6) and Section IV.D (Results, Table I discussion)]
"The deep-learning efficiency score (DES) metric, defined in Eq. 6, was used to identify efficient point cloud generation methods for deep learning. ... Meanwhile, the percentile 1% method attains the highest DES value of 0.48, while also demonstrating superior point generation ability while reducing data volume, making it well-suited for deep learning model training."

Eq. 6 defines DES as a density-normalized combination of min-max normalized PSNR and SSIM, with D(i) equal to point-cloud density. The paper then treats 'highest DES' as evidence that the 1% percentile method is 'well-suited for deep learning model training.' No downstream detector, segmentation experiment, or non-learning baseline validates DES as a proxy for deep-learning suitability; the conclusion is therefore equivalent to saying that 1% has the largest value of the quantity the authors chose to call the deep-learning efficiency score. In particular, the 1%-over-5% ranking is driven by dividing by D(i) (1.11 vs 4.46), not by any measured task performance, so the headline recommendation reduces to the metric's own construction.

full rationale

The tensor-generation model itself is not circular: 4DR P2T is trained against held-out K-Radar ground-truth tensors using cGAN, L1, and perceptual losses, and the reported PSNR/SSIM (30.39 dB / 0.96) are direct reconstruction metrics computed on held-out data. The percentile/CFAR comparison in Table I is likewise an honest empirical ranking on those metrics. The only circular step is the 'deep-learning suitability' conclusion: DES (Eq. 6) is defined by the authors as a density-normalized combination of min-max normalized PSNR and SSIM, and Section IV.D converts 'highest DES' directly into 'well-suited for deep learning model training.' Because no downstream detector, segmentation model, or non-learning tensor-voxelization baseline is used to validate DES as a proxy, the 1%-optimal recommendation is a restatement of the metric's definition (including the equal weighting alpha=beta=0.5 and density division D(i)), not an independently demonstrated result. Self-citations to K-Radar [10], RTNH+ [11], and Enhanced K-Radar [12] supply dataset and preprocessing choices; they are not used as a uniqueness theorem or to forbid alternatives, so they do not add circularity. Overall score 4: the core reconstruction is independent, but one of the paper's two headline conclusions reduces to a self-defined metric.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The model introduces no new physical entities. The central claim rests on four domain assumptions: the point-cloud-to-tensor decoding assumption, the validity of 2D projected metrics, the decision to drop Doppler, and the validity of the hand-weighted DES metric. Free parameters include the unreported loss weights, the arbitrary DES weights, and the ROI boundaries.

free parameters (3)

lambda_L1 and lambda_perc loss weights = not reported
Equation (1) defines the final loss as L_cGAN + lambda_L1 * L1 + lambda_perc * L_perc, but the values of lambda_L1 and lambda_perc are never stated; they affect reconstruction fidelity but are not tunable by the reader.
DES weights alpha and beta = alpha = 0.5, beta = 0.5
Section IV.C sets alpha = beta = 0.5 by hand. The conclusion that the 1% percentile is optimal for deep learning is a direct consequence of this choice; different weights could change the ranking.
ROI boundaries = x [0, 76.8], y [-16, 16], z [-2, 10.8]
Section IV.B selects the ROI based on the RTNH WIDE model; tensor quality scores are computed only inside this region and may not generalize to longer ranges or other sensor placements.

assumptions (4)

domain assumption A point cloud can be treated as an encoded version of the radar tensor, and the cGAN can learn to decode it back to the original tensor.
Introduced in Section I and used to justify using tensor generation performance as a proxy for point cloud quality.
domain assumption PSNR and SSIM computed on a mean-pooled 2D projection are valid proxies for 3D radar tensor fidelity.
Section IV.C evaluates 3D outputs only after height-axis mean pooling; this projection may hide errors along elevation.
domain assumption Excluding Doppler information from the tensor target does not compromise the evaluation of spatial synthesis quality.
Section III.A drops Doppler to reduce model complexity, citing RTNH [10] that detection performance is sufficient without it.
ad hoc to paper The DES metric with equal weights alpha = beta = 0.5 and min-max normalization yields a meaningful ranking of point cloud methods.
Section IV.C defines DES; the weights are chosen by hand, and the conclusion about the 1% percentile depends on this choice.

how reviews work

0 comments

Cite this review

Pith. "Pith review of 4DR P2T: 4D Radar Tensor Synthesis with Point Clouds." pith.science (2026). https://pith.science/paper/RO673GHC

@misc{pith2026250205550,
  author       = {Pith},
  title        = {Pith review of: 4DR P2T: 4D Radar Tensor Synthesis with Point Clouds},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RO673GHC}},
  note         = {Machine review of arXiv:2502.05550}
}

read the original abstract

In four-dimensional (4D) Radar-based point cloud generation, clutter removal is commonly performed using the constant false alarm rate (CFAR) algorithm. However, CFAR may not fully capture the spatial characteristics of objects. To address limitation, this paper proposes the 4D Radar Point-to-Tensor (4DR P2T) model, which generates tensor data suitable for deep learning applications while minimizing measurement loss. Our method employs a conditional generative adversarial network (cGAN), modified to effectively process 4D Radar point cloud data and generate tensor data. Experimental results on the K-Radar dataset validate the effectiveness of the 4DR P2T model, achieving an average PSNR of 30.39dB and SSIM of 0.96. Additionally, our analysis of different point cloud generation methods highlights that the 5% percentile method provides the best overall performance, while the 1% percentile method optimally balances data volume reduction and performance, making it well-suited for deep learning applications.

Figures

Figures reproduced from arXiv: 2502.05550 by the authors.

**Figure 2.** 4D Radar signal processing and data representation [2], [5], [6]. The Radar power values are normalized and represented using colors. The Radar [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Overall structure of 4DR P2T. The encoder utilizes 3D sparse convolution to process 4D Radar point cloud data, while the decoder employs 3D [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Qualitative experimental results of 4DR P2T. The top part shows the front camera image and LiDAR point cloud as reference data to understand [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 19 canonical work pages

[1]

An overview of radar clutter simulation

Ting An, Sai Zhu, Lidong Wang, and Zhiyong Liu. An overview of radar clutter simulation. In IOP conference series: materials science and engineering , volume 452, page 042198. IOP Publishing, 2018

work page 2018
[2]

4d millimeter-wave radar in autonomous driving: A survey

Zeyu Han, Jiahao Wang, Zikun Xu, Shuocheng Yang, Lei He, Shaob- ing Xu, Jianqiang Wang, and Keqiang Li. 4d millimeter-wave radar in autonomous driving: A survey. arXiv preprint arXiv:2306.04242 , 2023

arXiv 2023
[3]

Rpfa-net: A 4d radar pillar feature attention network for 3d object detection

Baowei Xu, Xinyu Zhang, Li Wang, Xiaomei Hu, Zhiwei Li, Shuyue Pan, Jun Li, and Yongqiang Deng. Rpfa-net: A 4d radar pillar feature attention network for 3d object detection. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC) , pages 3061–

work page 2021
[4]

Dpft: Dual perspective fusion transformer for camera-radar-based object detection

Felix Fent, Andras Palffy, and Holger Caesar. Dpft: Dual perspective fusion transformer for camera-radar-based object detection. arXiv preprint arXiv:2404.03015, 2024

arXiv 2024
[5]

Introduction to 4d radar: Hard- ware, mimo, signal processing, dataset, and ai

Dong-Hee Paek and Junfeng Guan. Introduction to 4d radar: Hard- ware, mimo, signal processing, dataset, and ai. In Tutorial on 2024 IEEE Intelligent V ehicles Symposium , 2024. Tutorial presentation

work page 2024
[6]

Exploring radar data representations in autonomous driving: A comprehensive review

S Yao et al. Exploring radar data representations in autonomous driving: A comprehensive review. arXiv preprint arXiv:2312.04861 , 2023

arXiv 2023
[7]

3drimr: 3d reconstruction and imaging via mmwave radar based on deep learning

Yue Sun, Zhuoming Huang, Honggang Zhang, Zhi Cao, and Deqiang Xu. 3drimr: 3d reconstruction and imaging via mmwave radar based on deep learning. In 2021 IEEE International Performance, Computing, and Communications Conference (IPCCC) , pages 1–8. IEEE, 2021

work page 2021
[8]

A novel radar point cloud generation method for robot environment perception

Yuwei Cheng, Jingran Su, Mengxin Jiang, and Yimin Liu. A novel radar point cloud generation method for robot environment perception. IEEE Transactions on Robotics , 38(6):3754–3773, 2022

work page 2022

Show all 25 references

[9]

U-net: Con- volutional networks for biomedical image segmentation

Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Con- volutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, par...

2015
[10]

K-radar: 4d radar object detection for autonomous driving in various weather conditions

Dong-Hee Paek, 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 , 35:3819–3829, 2022

2022
[11]

Rtnh+: Enhanced 4d radar object detection network using two-level pre- processing and vertical encoding

Seung-Hyun Kong, Dong-Hee Paek, and Sangyeong Lee. Rtnh+: Enhanced 4d radar object detection network using two-level pre- processing and vertical encoding. IEEE Transactions on Intelligent V ehicles, 2024

2024
[12]

En- hanced k-radar: Optimal density reduction to improve detection per- formance and accessibility of 4d radar tensor-based object detection

Dong-Hee Paek, Seung-Hyun Kong, and Kevin Tirta Wijaya. En- hanced k-radar: Optimal density reduction to improve detection per- formance and accessibility of 4d radar tensor-based object detection. In 2023 IEEE Intelligent V ehicles Symposium (IV) , pages 1–6. IEEE, 2023

2023
[13]

Raw high-definition radar for multi-task learning

Julien Rebut, Arthur Ouaknine, Waqas Malik, and Patrick P ´erez. Raw high-definition radar for multi-task learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 17021–17030, 2022

2022
[14]

Image-to-image translation with conditional adversarial networks

Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125–1134, 2017

2017
[15]

High-resolution image synthesis and semantic manipulation with conditional gans

Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 8798– 8807, 2018

2018
[16]

Radsimreal: Bridging the gap be- tween synthetic and real data in radar object detection with simulation

Oded Bialer and Yuval Haitman. Radsimreal: Bridging the gap be- tween synthetic and real data in radar object detection with simulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15407–15416, 2024

2024
[17]

L2r gan: Lidar-to-radar translation

Leichen Wang, Bastian Goldluecke, and Carsten Anklam. L2r gan: Lidar-to-radar translation. In Proceedings of the Asian Conference on Computer Vision, 2020

2020
[18]

V oxelnet: End-to-end learning for point cloud based 3d object detection

Yin Zhou and Oncel Tuzel. V oxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 4490– 4499, 2018

2018
[19]

Lidar-to-radar translation based on voxel feature extraction module for radar data augmentation

Jinho Lee, Geonkyu Bang, Takaya Shimizu, Masato Iehara, and Shunsuke Kamijo. Lidar-to-radar translation based on voxel feature extraction module for radar data augmentation. Sensors, 24(2):559, 2024

2024
[20]

Sparse convolutional neural networks

Baoyuan Liu, Min Wang, Hassan Foroosh, Marshall Tappen, and Mar- ianna Pensky. Sparse convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 806–814, 2015

2015
[21]

3d semantic segmentation with submanifold sparse convolutional networks

Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 9224–9232, 2018

2018
[22]

K- radar: 4d radar object detection for autonomous driving in various weather conditions

Dong-Hee Paek, Seung-Hyun Kong, and Kevin Tirta Wijaya. K- radar: 4d radar object detection for autonomous driving in various weather conditions. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track , 2022

2022
[23]

Adam: A method for stochastic optimization

Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 , 2014

2014 arXiv
[24]

Optimized transmission of jpeg2000 streams over wireless channels

Nikolaos Thomos, Nikolaos V Boulgouris, and Michael G Strintzis. Optimized transmission of jpeg2000 streams over wireless channels. IEEE Transactions on image processing , 15(1):54–67, 2005

2005
[25]

Robust transmission of jpeg2000 encoded images over packet loss channels

Xiangjun Li and Jianfei Cai. Robust transmission of jpeg2000 encoded images over packet loss channels. In 2007 IEEE International Conference on Multimedia and Expo , pages 947–950. IEEE, 2007

2007

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

An overview of radar clutter simulation

Ting An, Sai Zhu, Lidong Wang, and Zhiyong Liu. An overview of radar clutter simulation. In IOP conference series: materials science and engineering , volume 452, page 042198. IOP Publishing, 2018

work page 2018

[2] [2]

4d millimeter-wave radar in autonomous driving: A survey

Zeyu Han, Jiahao Wang, Zikun Xu, Shuocheng Yang, Lei He, Shaob- ing Xu, Jianqiang Wang, and Keqiang Li. 4d millimeter-wave radar in autonomous driving: A survey. arXiv preprint arXiv:2306.04242 , 2023

arXiv 2023

[3] [3]

Rpfa-net: A 4d radar pillar feature attention network for 3d object detection

Baowei Xu, Xinyu Zhang, Li Wang, Xiaomei Hu, Zhiwei Li, Shuyue Pan, Jun Li, and Yongqiang Deng. Rpfa-net: A 4d radar pillar feature attention network for 3d object detection. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC) , pages 3061–

work page 2021

[4] [4]

Dpft: Dual perspective fusion transformer for camera-radar-based object detection

Felix Fent, Andras Palffy, and Holger Caesar. Dpft: Dual perspective fusion transformer for camera-radar-based object detection. arXiv preprint arXiv:2404.03015, 2024

arXiv 2024

[5] [5]

Introduction to 4d radar: Hard- ware, mimo, signal processing, dataset, and ai

Dong-Hee Paek and Junfeng Guan. Introduction to 4d radar: Hard- ware, mimo, signal processing, dataset, and ai. In Tutorial on 2024 IEEE Intelligent V ehicles Symposium , 2024. Tutorial presentation

work page 2024

[6] [6]

Exploring radar data representations in autonomous driving: A comprehensive review

S Yao et al. Exploring radar data representations in autonomous driving: A comprehensive review. arXiv preprint arXiv:2312.04861 , 2023

arXiv 2023

[7] [7]

3drimr: 3d reconstruction and imaging via mmwave radar based on deep learning

Yue Sun, Zhuoming Huang, Honggang Zhang, Zhi Cao, and Deqiang Xu. 3drimr: 3d reconstruction and imaging via mmwave radar based on deep learning. In 2021 IEEE International Performance, Computing, and Communications Conference (IPCCC) , pages 1–8. IEEE, 2021

work page 2021

[8] [8]

A novel radar point cloud generation method for robot environment perception

Yuwei Cheng, Jingran Su, Mengxin Jiang, and Yimin Liu. A novel radar point cloud generation method for robot environment perception. IEEE Transactions on Robotics , 38(6):3754–3773, 2022

work page 2022