AdaRadar: Rate Adaptive Spectral Compression for Radar-based Perception

Jinho Park; Se Young Chun; Mingoo Seok

cs.CVMar 18, 2026

AdaRadar: Rate Adaptive Spectral Compression for Radar-based Perception

Authors: Jinho Park, Se Young Chun, Mingoo Seok

Organizations: Columbia University · Seoul National University

Abstract

Radar is a critical perception modality in autonomous driving systems due to its all-weather characteristics and ability to measure range and Doppler velocity. However, the sheer volume of high-dimensional raw radar data saturates the communication link to the computing engine (e.g., an NPU), which is often a low-bandwidth interface with data rate provisioned only for a few low-resolution range-Doppler frames. A generalized codec for utilizing high-dimensional radar data is notably absent, while existing image-domain approaches are unsuitable, as they typically operate at fixed compression ratios and fail to adapt to varying or adversarial conditions. In light of this, we propose radar data compression with adaptive feedback. It dynamically adjusts the compression ratio by performing gradient descent from the proxy gradient of detection confidence with respect to the compression rate. We employ a zeroth-order gradient approximation as it enables gradient computation even with non-differentiable core operations--pruning and quantization. This also avoids transmitting the gradient tensors over the band-limited link, which, if estimated, would be as large as the original radar data. In addition, we have found that radar feature maps are heavily concentrated on a few frequency components. Thus, we apply the discrete cosine transform to the radar data cubes and selectively prune out the coefficients effectively. We preserve the dynamic range of each radar patch through scaled quantization. Combining those techniques, our proposed online adaptive compression scheme achieves over 100x feature size reduction at minimal performance drop (~1%p). We validate our results on the RADIal, CARRADA, and Radatron datasets.

Figures & tables

Figure 1: Adaptive codec with task-guided feedback control. We propose an adaptive codec that compresses high-dimensional range-Doppler data by pruning spectral domain coefficients. A feedback loop adaptively regulates the compression ratio, guided by the performance of downstream perception tasks.

Figure 2: AdaRadar: Online rate-adaptive radar compression framework. Our proposed method introduces a feedback loop in which the proxy gradient is computed from the detection outputs to update the compression ratio adaptively. This avoids the need for backpropagation through the communication channel. The radar tensor is compressed using DCT, adaptive spectral pruning, and scaled quantization, then transmitted to the compute side. In an object-detection setting, the neural network produces detection results from decompressed radar data cubes. The loop uses proposed bounding boxes to estimate the proxy gradient, thereby updating the pruning rate.

Figure 3: Radar-based perception system overview. An FMCW radar transmits linearly swept‑frequency chirps. The incoming echo is mixed with a copy of the transmitted chirp at the receiver, yielding a de‑chirped intermediate‑frequency signal that the ADC digitizes to form a raw radar tensor. Successive FFTs along the fast‑time and slow‑time axes convert this tensor into a range–Doppler cube. The large, raw radar tensor is transferred to the NPU over the power-hungry sensor-to-compute link for network inference.

Figure 4: Motivation for spectral pruning and quantization. (a) The DCT coefficient magnitudes are clustered in the high-frequency bins. (b) Their histogram is sharply peaked, highlighting strong sparsity and clear opportunities for compression.

Figure 5: Motivation for surrogate objective choice. (a) AP and (b) AR vs. confidence score yield a high correlation.

Method	Bit	Prune	Bit rate [bpp] $\downarrow$	Detection			Segmentation
Method	Bit	ratio $\uparrow$	(Comp. ratio $\uparrow$ )	P (%) $\uparrow$	R (%) $\uparrow$	F 1 (%) $\uparrow$	mIoU (%) $\uparrow$
Uncompressed baseline [ 3 ]	32	-	32 (1×)	97.24	95.93	96.58	75.97
Index-value [ 9 ]	32	12×	2.67 (12×)	97.55	62.12	75.91	49.86
Adaptive compression (Ours)	4	12.57×	0.32 ( 101× )	96.25	94.04	95.13	79.34

Table 1: Compression on RADIal. We compare the detection and segmentation performance of adaptive rate control with prior work on the RADIal test set using FFTRadNet as the backbone. Our adaptive rate control achieves 101× compression on average while maintaining the performance close to the baseline. ‘P’ and ‘R’ respectively stand for precision and recall. We highlight in bold the best-performing result among methods other than the baseline.

Method	Bit	PR $\uparrow$	BR [bpp] $\downarrow$	CR $\uparrow$	mIoU (%) $\uparrow$	mDice (%) $\uparrow$
Uncompressed baseline [ 48 ]	32	-	32	1×	55.25	67.13
Index-value [ 9 ]	32	29×	1.10	29×	38.96	46.90
Adaptive compression (Ours)	8	29.26×	0.27	117×	54.03	65.87

Table 2: Compression on CARRADA. We compare our adaptive compression strategy with prior arts on mIoU and mDice metrics on the CARRADA dataset using the TMVA-Net. Our adaptive method maintains segmentation performance close to the baseline despite a large reduction (117×) in data rate. ‘PR’, ‘BR’, and ‘CR’ respectively stand for the prune ratio, bit rate, and total compression ratio. Bold text indicates the top-performing results, excluding the baseline.

Method	Bit	PR $\uparrow$	BR [bpp] $\downarrow$	CR $\uparrow$	mAP (%) $\uparrow$	AP 50 (%) $\uparrow$	AP 75 (%) $\uparrow$
Baseline [ 11 ]	32	-	32	1×	46.07	83.60	44.16
+Index-value [ 9 ]	32	7.5×	4.27	7.5×	45.72	80.44	47.54
+Adaptive (Ours)	8	7.5×	1.07	30×	48.46	83.69	49.07

Table 3: Compression on Radatron. We compare compression strategies on mAP, AP 50 , and AP 75 while reducing the bit rate. Our adaptive method achieves higher detection performance than both the baseline and index-value compression while operating at a much lower bit rate (30

\times

compression). ‘PR’, ‘BR’, and ‘CR’ respectively stand for the prune ratio, bit rate, and total compression ratio. The best-performing result, except for the baseline method, is highlighted in bold.

Figure 6: Online adaptation. Bit rate time series (above). Scene-wise average recall change (below).

Figure 7: Rate-accuracy tradeoff. (a) Average precision and (b) average recall vs. bit rate measured with a static compression rate on the RADIal test set with

4

-bit quantization.

Figure 8: Qualitative comparisons. Networks are trained on raw radar tensors, visualized as range-Doppler (RD) images, along with ground-truth labels for freeway segmentation and vehicle detection. Camera images only provide contextual reference for the scene. RD images map Range to the $y$ -axis and Doppler to the $x$ -axis. We highlight the detected car in red and the segmented map in green in the bird’s-eye-view visualization. Notably, the compressed RD outperforms the uncompressed baseline.

Figure 9: Adaptive spectral pruning vs. index-value-based compression. Adaptive spectral pruning-based compression outperforms index-value-based compression [ 9 ] with much more resiliency to error. Both methods utilize Float32 , and are tested on the RADIal test set.

Figure S1 : Compression, decompression, communication cost analysis. Our compression and decompression processes significantly reduce latency compared to the baseline system, which transmits the original radar tensor.

Surrogate objective	Corr. w/ AP	Corr. w/ AR
Confidence (Ours)	0.841	0.857
Range stdev.	-0.129	-0.045
Azimuth stdev.	-0.070	-0.016

Table S1: Correlation between OD prediction output and performance metrics. The performance metrics, such as AP and AR, are highly correlated with confidence but not with other objectives.

Surrogate objective	CR $\uparrow$	CR stdev. $\downarrow$	Precision $\uparrow$	Recall $\uparrow$	F 1 $\uparrow$	mIoU $\uparrow$
Confidence (Ours)	6.58	0.91	0.969	0.948	0.958	0.796
Range stdev.	60.9	22.1	0.941	0.785	0.856	0.721
Azimuth stdev.	87.0	9.8	0.930	0.762	0.837	0.708
Random	34.1	29.7	0.945	0.834	0.886	0.741

Table S2: Surrogate objective ablation with range standard deviation, azimuth standard deviation, and a random variable. We observe that different choices of surrogate objective do not correlate well with the downstream task performance, leading to severe degradation in the perception performance. In those cases, such a significant drop in F 1 score bears similar performance to the random variable.

$\lambda$	PR $\uparrow$	BR [bpp] $\downarrow$	CR $\uparrow$	Precision (%) $\uparrow$	Recall (%) $\uparrow$	F 1 (%) $\uparrow$	mIoU (%) $\uparrow$
1	12.57	0.32	100.56	96.25	94.03	95.13	79.34
3.3	13.83	0.29	110.68	96.21	93.87	95.03	79.21
10	16.09	0.25	128.74	96.07	93.64	94.84	79.12
33	20.75	0.19	166.03	95.70	93.07	94.36	78.87
100	27.91	0.14	223.29	95.48	91.66	93.53	78.18

Table S3 : Hyperparameter ablation with

\lambda

. We analyze the trade-off between bit rate, compression ratio, and downstream performance on detection and segmentation through modulation of

\lambda

. ‘PR’, ‘BR’, and ‘CR’ stand for prune ratio, bit rate, and compression ratio, respectively. We use 4-bit quantization with a block length of

M=64

. The pruning ratio is initially set to 12, and the learning rate is configured as

\eta=1

Compression	Condition	Noise Scale	Comp. Ratio $\uparrow$	Prun. Ratio $\uparrow$	Precision $\uparrow$	Recall $\uparrow$	F 1 $\uparrow$	mIoU $\uparrow$
✗ No	Mild	1	1.0	1.0	0.973	0.922	0.947	0.774
✓ Yes	Mild	1	36.8	4.6	0.969	0.910	0.939	0.777
✗ No	Moderate	2	1.0	1.0	0.957	0.884	0.919	0.694
✓ Yes	Moderate	2	33.7	4.2	0.967	0.860	0.910	0.692
✗ No	Severe	5	1.0	1.0	0.931	0.771	0.844	0.618
✓ Yes	Severe	5	25.6	3.2	0.939	0.720	0.815	0.602

Table S4: Proxy gradient stability at varying additive Gaussian noise levels on the RADIal test set. We test the stability of our feedback loop by adding Gaussian noise to the radar input (RD map). Even under the noisy scenario, our adaptive control-based compression demonstrates similar performance to the non-compressed baseline.

Methods	Hyperparameters	BR [bpp] $\downarrow$	CR $\uparrow$	Precision (%) $\uparrow$	Recall (%) $\uparrow$	F 1 (%) $\uparrow$	mIoU (%) $\uparrow$
Baseline	-	32	1×	97.24	95.93	96.58	75.97
Ours	$s_{\texttt{FxP}}=4\text{-bit}$	0.32	100.5×	96.25	94.04	95.13	79.34
AE	$\lambda_{ae}=1.0\cdot 10^{-3}$	1.66	19.2×	97.20	95.67	96.43	75.45
AE	$\lambda_{ae}=3.3\cdot 10^{-4}$	1.07	29.75×	97.22	94.85	96.02	74.42
AE	$\lambda_{ae}=1.0\cdot 10^{-4}$	0.59	54.1×	97.40	92.82	95.05	70.80
AE	$\lambda_{ae}=3.3\cdot 10^{-5}$	0.27	117.1×	97.49	77.05	86.08	61.04

Table S5: Benchmark with extended baselines. We compare radar compression with autoencoders, H.264, Gumbel-sigmoid-based pruning, and CFAR on RADIal.

Figure S2 : Difference in distribution between image and radar tensors . This illustrates that while natural (a) images span a broad dynamic range, (b) radar signals exhibit extreme sparsity and a heavy-tailed profile (1st/99th percentiles at

\pm

1.74), necessitating specialized compression beyond standard image-domain codecs.

Methods	Bit $\downarrow$	PR $\uparrow$	BR [bpp] $\downarrow$	CR $\uparrow$	Precision (%) $\uparrow$	Recall (%) $\uparrow$	F 1 (%) $\uparrow$	mIoU (%) $\uparrow$
Baseline [ 55 ]	32	1	32	1×	90.61	93.67	92.11	80.56
Ours	4	2	2	16×	90.94	93.19	92.05	80.34
Ours	4	4	1	32×	90.55	91.89	91.22	79.56

Table S6: Compression after learned FFT. We demonstrate the effectiveness of AdaRadar even with learned FFT. We use T-FFTRadNet for detection and segmentation using the RADIal test set. Our compression method achieves 32× data reduction with negligible performance loss. (PR: prune rate; BR: bit rate; CR: compression rate.)

Figure S3 : Qualitative visualization on range-Doppler map. (a) Original and (b) reconstructed RD map for a single channel with 8-bit quantization and a pruning ratio of 1. (c)-(f) Magnified 64×64 patches at pruning ratios of $\{1,5,10,20\}$ , respectively. (g) SNR vs. bit rate trade-off where numerical values are listed in Table S7 .

Comp. Ratio $\uparrow$	SNR [dB] $\uparrow$	RAE mean $\downarrow$	RAE max $\downarrow$
1	144.17	0.000	0.000
5	35.45	0.010	0.329
10	32.26	0.017	0.378
15	30.77	0.022	0.389
20	29.75	0.025	0.390
30	28.36	0.031	0.383

Table S7: SNR & RAE against compression. Mean and maximum relative absolute errors (RAE) at different compression ratios.

Panel of Figure 23 — Figure S4 : DCT coefficient magnitudes. (a) CARRADA and (b) Radatron datasets. Unlike the RADIal dataset, the spectral bins are more focused on low-frequency bins. Our adaptive pruning is agnostic to such spectral structures, since it ranks coefficients by energy.

Figure S5 : Effect of quantization bit width on RADIal. (a) Average precision, (b) average recall, and (c) F 1 score against the bit rate modulation from pruning. We observe that quantization up to 4 bits does not affect the performance compared to that of 8-bit and 16-bit. We use the FFTRadNet on the RADIal dataset without any fine-tuning with the block length

M=64

for all cases.

Figure S6 : Effect of quantization bit width on CARRADA. (a) mIoU and (b) mDice reported versus the bit rate reduction. We observe that quantization up to 8 bits does not affect the performance compared to that of 16-bit. We use the TMVA-Net on the CARRADA dataset without any fine-tuning with the block length

M=16

for all cases.

Figure S7 : Effect of quantization bit width on Radatron. (a) mAP, (b) AP 50 , and (c) AP 75 vs. the bit rate. We observe that quantization up to 8 bits does not affect the performance compared to that of 4-bit and 16-bit. We use the baseline Radatron model on the Radatron dataset without any fine-tuning with the block length

M=16

for all cases.

Figure S8 : Effect of block length on RADIal. (a) Average precision, (b) average recall, and (c) F 1 score against the block length modulation. We observe that compression with a larger block length generally performs better. We use the FFTRadNet on the RADIal dataset without any fine-tuning with 4-bit quantization for all cases.

Figure S9 : Effect of block length on CARRADA. (a) mIoU and (b) mDice reported versus the change in block length. We observe that a larger block is preferred for a higher pruning ratio and a smaller block size for a lower pruning ratio. We use the TMVA-Net on the CARRADA dataset without any fine-tuning with 8-bit quantization for all cases.

Figure S10 : Effect of block length on Radatron. (a) mAP, (b) AP 50 , and (c) AP 75 vs. block size change. We observe that a larger patch performs better for higher pruning ratios. We use the baseline Radatron model on the Radatron dataset without any fine-tuning with 8-bit quantization for all cases.

Figure S11 : Additional qualitative comparisons. Camera images only provide contextual reference for the scene. RD images map Range to the $y$ -axis and Doppler to the $x$ -axis. We highlight the detected car in red and the segmented map in green in the bird’s-eye-view visualization.

Autonomous Driving Perception

Autonomous driving perception enables vehicles to detect and interpret road users, drivable space, and hazards using onboard and shared sensors.

See all 379 papers

Feb 23, 2026

RADE-Net: Robust Attention Network for Radar-Only Object Detection in Adverse Weather

Automotive perception systems are obligated to meet high requirements. While optical sensors such as Camera and Lidar struggle in adverse weather conditions, Radar provides a more robust perception performance, effectively penetrating fog, rain, and snow. Since full Radar tensors have large data sizes and very few datasets provide them, most Radar-based approaches work with sparse point clouds or 2D projections, which can result in information loss. Additionally, deep learning methods show potential to extract richer and more dense features from low level Radar data and therefore significantly increase the perception performance. Therefore, we propose a 3D projection method for fast-Fourier-transformed 4D Range-Azimuth-Doppler-Elevation (RADE) tensors. Our method preserves rich Doppler and Elevation features while reducing the required data size for a single frame by 91.9% compared to a full tensor, thus achieving higher training and inference speed as well as lower model complexity. We introduce RADE-Net, a lightweight model tailored to 3D projections of the RADE tensor. The backbone enables exploitation of low-level and high-level cues of Radar tensors with spatial and channel-attention. The decoupled detection heads predict object center-points directly in the Range-Azimuth domain and regress rotated 3D bounding boxes from rich feature maps in the cartesian scene. We evaluate the model on scenes with multiple different road users and under various weather conditions on the large-scale K-Radar dataset and achieve a 16.7% improvement compared to their baseline, as well as 6.5% improvement over current Radar-only models. Additionally, we outperform several Lidar approaches in scenarios with adverse weather conditions. The code is available under https://github.com/chr-is-tof/RADE-Net.

Christof Leitgeb, Thomas Puchleitner, Max Peter Ronecker +1

Infineon Technologies AG, Institute of Visual Computing, Graz University of Technology, Graz, Austria · Infineon Technologies AG, Graz, Austria · Virtual Vehicle Research GmbH, Institute of Visual Computing, Graz University of Technology, Graz, Austria +1

3D Object Detection mmWave Radar Sensing

Ask PDF

Mar 24, 2026

LiZIP: An Auto-Regressive Compression Framework for LiDAR Point Clouds

The massive volume of data generated by LiDAR sensors in autonomous vehicles creates a bottleneck for real-time processing and vehicle-to-everything (V2X) transmission. Existing lossless compression methods often force a trade-off: industry standard algorithms (e.g., LASzip) lack adaptability, while deep learning approaches suffer from prohibitive computational costs. This paper proposes LiZIP, a lightweight, near-lossless zero-drift compression framework based on neural predictive coding. By utilizing a compact Multi-Layer Perceptron (MLP) to predict point coordinates from local context, LiZIP efficiently encodes only the sparse residuals. We evaluate LiZIP on the NuScenes and Argoverse datasets, benchmarking against GZip, LASzip, and Google Draco (configured with 24-bit quantization to serve as a high-precision geometric baseline). Results demonstrate that LiZIP consistently achieves superior compression ratios across varying environments. The proposed system achieves a 7.5%-14.8% reduction in file size compared to the industry-standard LASzip and outperforms Google Draco by 8.8%-11.3% across diverse datasets. Furthermore, the system demonstrates generalization capabilities on the unseen Argoverse dataset without retraining. Against the general purpose GZip algorithm, LiZIP achieves a reduction of 38%-48%. This efficiency offers a distinct advantage for bandwidth constrained V2X applications and large scale cloud archival.

Aditya Shibu, Kayvan Karim, Claudio Zito

MACS, Heriot Watt University, Dubai, United Arab Emirates

Point Cloud Compression Predictive Coding

Ask PDF

Sep 16, 2026

Accuracy- and Real-Time-Aware 4D Radar Preprocessing for Autonomous Driving Perception Systems

4D radar has emerged as a promising next-generation sensor for improving the robustness of autonomous driving perception systems because of its stable sensing capability under adverse weather conditions. However, deploying 4D radar in embedded environments with limited hardware resources requires radar-representation preprocessing that jointly considers perception accuracy, real-time performance, and computational complexity. This paper proposes a preprocessing framework for 4D-radar-based 3D object detection. First, Percentile-based 3D Shape Preservation (P3DP) extracts point clouds from radar tensors while preserving object-shape information and suppressing noise and false alarms. Second, Multi-frame-based Noise Point Discrimination using Kernel Density Estimation (MF-KDE) improves the density and reliability of sparse radar point clouds. Finally, Embedded & NetScore (ENS) evaluates suitability for embedded deployment by jointly considering accuracy, real-time performance, adverse-weather robustness, and model complexity.

Woo-Jin Jung, Dong-Hee Paek, Jeong-Su Park +1

Cho Chun Shik Graduate School of Mobility Korea Advanced Institute of Science and Technology (KAIST) Daejeon 34051, Republic of Korea · Mechanical Engineering Research Institute Korea Advanced Institute of Science and Technology (KAIST) Daejeon 34141, Republic of Korea · Autonomous Driving Perception Technology Vanguard Team Hyundai Motor Company Seongnam-si 13529, Republic of Korea

Autonomous Driving Perception 3D Object Detection Real-Time Object Detection Millimeter-Wave Radar

Ask PDF

Abstract

Autonomous Driving Perception

Explore similar work

RADE-Net: Robust Attention Network for Radar-Only Object Detection in Adverse Weather

LiZIP: An Auto-Regressive Compression Framework for LiDAR Point Clouds

Accuracy- and Real-Time-Aware 4D Radar Preprocessing for Autonomous Driving Perception Systems