Pith. sign in

REVIEW 3 major objections 6 minor 24 references

Multi-Representation Adapter with Neural Architecture Search for Efficient Range-Doppler Radar Object Detection

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · 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 Range-Doppler radar detector sees more when the same map is fed twice—as a heatmap and a grayscale image—through a searchable adapter branch, reaching state-of-the-art mAP@50 under 3M parameters on RADDet and CARRAD

desk verdict Worth a serious look for the RD-radar-detection crowd, but the headline SOTA margin over DAROD on RADDet is not yet trustworthy because the authors made their own RD spectrograms without showing they match what the baselines consumed. read the letter →

arxiv 2509.01280 v1 pith:7AL7OJZJ submitted 2025-09-01 cs.CV

Zhiwei Lin , Weicheng Zheng , Yongtao Wang This is my paper
classification cs.CV
keywords Range-Dopplerradarobjectdetectionmulti-representationadapterbranchneuralarchitecturesearchYOLOheatmapgrayscale
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 sets out to show that a Range-Doppler radar map, usually treated as a single image, contains more detectable information when a detector sees two representations of it at once: a pseudocolor heatmap that emphasizes object-level structure and a grayscale image that preserves raw intensity texture. The authors build a two-branch detector on a YOLO-type base, with an adapter branch for the grayscale input and a sequence of exchanger modules that pass features in both directions, fused by a channel-weighted blending operation. They then wrap the adapter in a weight-sharing supernet and use one-shot neural architecture search to choose per-layer widths and fusion choices. On the RADDet and CARRADA benchmarks they report the best accuracy-efficiency trade-off to date, including a large mAP@50 lead over prior RD detectors on RADDet with only 2.65M parameters. The case matters because radar sensors work in darkness, fog, and rain, and an efficient RD-only detector could complement or replace cameras on cost- and power-constrained vehicles.

What carries the argument

The mechanism that carries the argument is a two-branch detector with an Adapter branch attached to a YOLO-type backbone. The backbone sees the heatmap representation; the Adapter sees the grayscale representation. Between them, Exchanger Modules alternate between Mode 1 (backbone features are primary, grayscale features auxiliary) and Mode 2 (roles reversed), so information flows in both directions. Each Exchanger's Primary-Auxiliary Fusion Module computes a channel-level importance weight from the auxiliary feature's global average pooling, then blends it into the primary feature; a lightweight coordinate-attention module and a learnable residual weight refine the result. Around this fixed

What would settle it

Take the exact RD spectrograms produced by the authors' ADC-to-RD pipeline and retrain the two strongest baselines (DAROD and FCOS+RiCL) on them without any architectural change. If either baseline reaches mAP@50 near 71.9 on RADDet, the central claim that the adapter design is responsible for the state-of-the-art result would be refuted; if the gap persists, the architecture claim survives.

Watch

Extended reading notes

Core claim

The central claim is that the standard practice of treating a Range-Doppler radar map as a single image leaves information on the table. Presenting the same RD map in two representations—a pseudocolor heatmap that brings out object-level structure and a grayscale intensity image that preserves fine texture—and running them through two branches that continually exchange features lets a small YOLO-type detector outperform much larger and more complex radar detectors. The exchange point is a set of Exchanger Modules alternating between two modes: one injects grayscale detail into the heatmap branch, the other injects heatmap context back into the grayscale branch, with a channel-weighted fusion

Load-bearing premise

The reported margin over prior work assumes the RD spectrograms the authors generate from RADDet raw ADC data are equivalent to the RD maps those earlier methods were trained and tested on; if preprocessing (FFT size, windowing, normalization) differs, part of the gap may come from the data format rather than from the proposed architecture.

Editorial extensions

If this is right

  • Reported results imply radar-only object detection can be both accurate and lightweight: the best subnet uses less than 3M parameters and still reports the highest mAP@50 on both benchmarks.
  • The multi-representation scheme—heatmap for high-level features, grayscale for texture—lets a relatively small YOLO-type backbone keep up with or beat heavier detectors.
  • The one-shot supernet search produces subnets that improve on the fixed adapter hand-configuration while cutting parameters from 3.24M to 2.65M on RADDet.
  • The Adapter branch is not tied to one backbone: integrating it into three YOLO variants raises mAP@50-95 on RADDet by 2-3 points with under 0.23M added parameters.

Reading between the lines

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

  • A natural follow-up is to hold one preprocessing pipeline fixed and vary only the architecture; the residual gap would show how much of the reported 25.3-point RADDet lead comes from the adapter design rather than from the authors' ADC-to-RD preprocessing.
  • Because the searched subnets can choose different fusion operations at different depths, inspecting the winning configuration could yield a simple hand-designed rule for where channel-gated fusion matters, making NAS unnecessary at deployment.
  • The adapter's gains on three YOLO variants suggest it acts as a general detail-injection module; a plausible transfer test is to apply the same two-branch, two-representation scheme to Range-Angle maps or to other sparse sensor grids such as LiDAR or ultrasound.
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.

Referee Report

3 major / 6 minor

Summary. The paper proposes an efficient radar object detection model for Range-Doppler (RD) maps. It uses two input representations of the same RD map — a heatmap fed to a YOLOv8-type backbone and a grayscale image fed to a newly designed Adapter branch. The Adapter branch contains a Stem, alternating Exchanger Modules with two modes, and a Primary-Auxiliary Fusion Module for bidirectional feature exchange and fusion. The model is further embedded in a weight-sharing supernet with elastic width and variable fusion operations, and a One-Shot NAS with evolutionary search is used to find a compact subnet. Experiments on RADDet and CARRADA report mAP@50 of 71.9 and 57.1, respectively, and the authors claim state-of-the-art performance. Internal ablations (Tables 4–7) show consistent improvements from the Adapter components across several YOLO variants.

Significance. If the SOTA comparisons are valid, the paper offers a practical accuracy/efficiency trade-off for radar RD detection and a sensible use of NAS over a multi-branch architecture. The internal evidence is mostly well designed: the component ablations (Tables 4–7) isolate the contributions of the Stem, both Exchanger modes, and the fusion option, and the generalization experiments across YOLOv8n/YOLOv9t/YOLOv10n give useful evidence that the Adapter is not a one-off trick. The reported parameter counts and Pareto-style comparisons are also valuable. However, the central 'new state-of-the-art' claim currently rests on a comparison that is confounded by the authors' own RD preprocessing pipeline for RADDet, and by the fact that the plain YOLOv8n baseline already exceeds the cited DAROD result by a large margin. Thus the significance of the headline claim is not yet established, even though the architectural contribution appears sound.

major comments (3)
  1. [§4.1, Table 2] The RADDet SOTA comparison is not yet supported. Section 4.1 states that the authors 'extract the raw analog-to-digital-converter data and process it using range and Doppler FFT to generate RD spectrograms', but no FFT length, windowing, normalization, dynamic range, or train/val split details are given. The baseline numbers for DAROD and FCOS+RiCL are taken from prior papers, not re-run on these generated inputs. This is load-bearing: Table 3 shows that a plain YOLOv8n with heatmap input already reaches 65.0 mAP@50 on the authors' RADDet pipeline, which is +18.4 over the cited DAROD result (46.6). The reported +25.3 margin for the full model is therefore mostly present before the Adapter/NAS contribution. The authors should document the preprocessing in full and either re-run the baselines on the same RD maps or provide official RD maps, so the comparison is apples-to-apples.
  2. [Table 2, CARRADA row] The CARRADA SOTA claim is weakened by the mAP@30 result. The proposed model achieves 62.1 mAP@30 versus DAROD's 70.7, a deficit of 8.6 points, while the mAP@50 gain is only +1.3. The text says the model shows 'favorable detection performance on multiple key metrics', which is not true for mAP@30. The authors should either explain this trade-off (e.g., better localization at looser IoU but worse at stricter localisation) or temper the claim that the model is uniformly better than DAROD on CARRADA.
  3. [§3.4, §4.1] The search/evaluation protocol introduces a selection effect that is not quantified. Section 3.4 says the top-5 subnets are fully trained and the best one is selected, and Section 4.1 gives the evolutionary search hyperparameters (P=50, T=20, k=15). Since no separate validation/test split details are given, and no variance or performance of the other four fully trained subnets is reported, the gains of YOLOv8n-Adapter-Search over YOLOv8n-Adapter may reflect, in part, selection over five trained models. Reporting the mean/min/max of the five subnets and the split used for selection versus final evaluation would make the efficiency claim more robust.
minor comments (6)
  1. [§4.1] NMS IoU threshold is set to 0.1, which is unusually low and can materially affect mAP. A sentence explaining why this follows previous works, or a sensitivity test, would help.
  2. [§4.2] Typo: 'supassing' should be 'surpassing'.
  3. [Keywords] 'ObjectionDetection' should be 'Object Detection'.
  4. [Table 1] The search space size is given as '220×37' but the table lists more than 5 choice blocks; please clarify how this number is computed and whether it is the product of all width/fusion options.
  5. [General] No code or preprocessing script is mentioned. Releasing the RD-generation script for RADDet would address the main reproducibility concern and strengthen the paper.
  6. [Figure 2 / Table 3] In the manuscript version I received, Figure 2 and the caption of Table 3 contain garbled symbol sequences, likely a font/encoding issue. Please ensure the camera-ready version renders all math and table text correctly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical architecture/NAS paper; reported accuracy is measured, not derived from inputs.

full rationale

The paper makes no first-principles derivation claim: its contributions are a dual-branch adapter architecture and a one-shot NAS procedure, evaluated by standard mAP on two datasets. I traced every reported performance claim to experimental measurement rather than to a definition or fitted parameter. Eq. (3) defines the search objective as validation accuracy of weight-shared subnets; the reported result is obtained after fully training the top-5 subnets and selecting the best, which is model selection, not fitting the metric into the metric. The inclusion of the proposed fusion module (Option 1) in the search space is not circular because the ablation in Table 6 independently compares the three fusion options. The only notable threat, that RADDet baselines DAROD/RiCL may have used different RD preprocessing than the authors' FFT-generated spectrograms, is a cross-paper comparability/soundness issue, not a circularity issue; it does not make the authors' mAP equal to an input by construction. There is no load-bearing self-citation, no imported uniqueness theorem, and no ansatz smuggled in via citation. Therefore, no significant circularity is present.

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

This paper does not derive a theory; it proposes an architecture. The central claim rests on the comparability of benchmark evaluations and on standard assumptions about training and NAS. No physical or unobservable entities are postulated; the modules are the contribution and are evaluated with internal ablations. The main free parameters are the preprocessing and evaluation choices that are not fully specified.

free parameters (4)
  • RADDet RD spectrogram preprocessing (FFT size, window, normalization) = unspecified
    Affects all RADDet numbers; the authors generate spectrograms from raw ADC data without reporting parameters, so the large gap over DAROD may be a preprocessing artifact.
  • NMS IoU threshold = 0.1
    Used at inference, follows DAROD but is an unusually permissive threshold that inflates mAP@50; affects reported numbers.
  • NAS search budget (population, iterations, top-k) = (50, 20, 15)
    Hand-chosen evolutionary search parameters that determine the searched subnet; no sensitivity analysis is provided.
  • Best-of-5 subnet selection = best of 5 fully trained subnets
    Reported result is the maximum over five fully trained subnets; this selection can inflate scores and no variance is reported.
assumptions (3)
  • domain assumption The RD spectrograms generated from RADDet raw ADC data are directly comparable to the RD maps used by DAROD and RiCL.
    Section 4.1 states the authors generate RD spectrograms with range and Doppler FFT; Table 2 compares mAP with published baselines without re-running them under the same preprocessing.
  • domain assumption Weight-sharing supernet accuracy is a reliable proxy for fully-trained subnet accuracy.
    Section 3.4 selects subnets using inherited weights and only fully trains the top 5; if the proxy is noisy, the chosen subnet may not be the true best.
  • domain assumption YOLOv8n is a strong and appropriate baseline for isolating the adapter's contribution.
    The adapter's measured gains (Tables 4-7) are relative to YOLOv8n; a different baseline could show different gains.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Representation Adapter with Neural Architecture Search for Efficient Range-Doppler Radar Object Detection." pith.science (2026). https://pith.science/paper/7AL7OJZJ

@misc{pith2026250901280,
  author       = {Pith},
  title        = {Pith review of: Multi-Representation Adapter with Neural Architecture Search for Efficient Range-Doppler Radar Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7AL7OJZJ}},
  note         = {Machine review of arXiv:2509.01280}
}
read the original abstract

Detecting objects efficiently from radar sensors has recently become a popular trend due to their robustness against adverse lighting and weather conditions compared with cameras. This paper presents an efficient object detection model for Range-Doppler (RD) radar maps. Specifically, we first represent RD radar maps with multi-representation, i.e., heatmaps and grayscale images, to gather high-level object and fine-grained texture features. Then, we design an additional Adapter branch, an Exchanger Module with two modes, and a Primary-Auxiliary Fusion Module to effectively extract, exchange, and fuse features from the multi-representation inputs, respectively. Furthermore, we construct a supernet with various width and fusion operations in the Adapter branch for the proposed model and employ a One-Shot Neural Architecture Search method to further improve the model's efficiency while maintaining high performance. Experimental results demonstrate that our model obtains favorable accuracy and efficiency trade-off. Moreover, we achieve new state-of-the-art performance on RADDet and CARRADA datasets with mAP@50 of 71.9 and 57.1, respectively.

Figures

Figures reproduced from arXiv: 2509.01280 by the authors.

Figure 1
Figure 1. The overall pipeline of YOLOv8n-Adapter. (a) The searchable YOLOv8 backbone branch takes the heatmap representation of RD maps as inputs. (b) The proposed Adapter branch uses the grayscale representation of RD maps as inputs. Exchanger Modules, alternating between two modes, establish a symmetric information flow, enhancing the overall feature extraction ability from RD maps. 2.3 Neural Architecture Search NAS aims … view at source ↗
Figure 2
Figure 2. The key components of the Adapter branch. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. Elastic convolution kernel based on weight sharing. We utilize dynamic channel pruning to obtain the selected weights. shared weight W for all subnets. The training of the supernet can be formulated as follows: WA = arg min W Ltrain(N (A, W)), (2) where Ltrain(·) is the loss function on the training set. For our model, the constructed supernet consists of multiple Choice Blocks, as shown in [PITH_FULL_IMAGE:figures… view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Model parameters, GFLOPs, latency vs. accuracy on RADDet dataset. Our method achieves the best accuracy-efficiency trade-offs. mAP@50, respectively. For the CARRADA dataset, our model achieves higher mAP@50, outperforming DAROD by 1.3 mAP@50. Our best subnet demonstrat…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 17 canonical work pages

  1. [1]

    In: ICCV (2021)

    Chen, B., Li, P., Li, B., Lin, C., Li, C., Sun, M., Yan, J., Ouyang, W.: Bn-nas: Neural architecture search with batch normalization. In: ICCV (2021)

    work page 2021

  2. [2]

    In: IEEE IV (2022)

    Decourt, C., VanRullen, R., Salle, D., Oberlin, T.: Darod: A deep automotive radar object detector on range-doppler maps. In: IEEE IV (2022)

    work page 2022

  3. [3]

    Leveraging Self-Supervised Instance Contrastive Learning for Radar Object Detection

    Decourt, C., VanRullen, R., Salle, D., Oberlin, T.: Leveraging self-supervised instance contrastive learning for radar object detection. arXiv preprint arXiv:2402.08427 (2024)

    work page Pith review arXiv 2024

  4. [4]

    arXiv preprint arXiv:2107.08430 (2021)

    Ge, Z.: Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430 (2021)

    arXiv 2021

  5. [5]

    In: ECCV (2020)

    Guo, Z., Zhang, X., Mu, H., Heng, W., Liu, Z., Wei, Y., Sun, J.: Single path one-shot neural architecture search with uniform sampling. In: ECCV (2020)

    work page 2020

  6. [6]

    In: WACV (2024)

    Haitman, Y., Bialer, O.: Boostrad: Enhancing object detection by boosting radar reflections. In: WACV (2024)

    work page 2024

  7. [7]

    In: CVPR (2021)

    Hou, Q., Zhou, D., Feng, J.: Coordinate attention for efficient mobile network design. In: CVPR (2021)

    work page 2021

  8. [8]

    https://doi.org/10.5281/zenodo.3908559, https://github.com/ultralytics/yolov5

    Jocher, G.: Ultralytics yolov5 (2020). https://doi.org/10.5281/zenodo.3908559, https://github.com/ultralytics/yolov5

    doi:10.5281/zenodo.3908559 2020

Show all 24 references
  1. [9]

    Jocher, G., Chaurasia, A., Qiu, J.: Ultralytics YOLO (Jan 2023), https://github.com/ultralytics/ultralytics

    2023

  2. [10]

    Jocher, G., Qiu, J.: Ultralytics yolo11 (2024), https://github.com/ultralytics/ultralytics

    2024

  3. [11]

    arXiv preprint arXiv:2209.02976 (2022)

    Li, C., Li, L., Jiang, H., Weng, K., Geng, Y., Li, L., Ke, Z., Li, Q., Cheng, M., Nie, W., et al.: Yolov6: A single-stage object detection framework for industrial applications. arXiv preprint arXiv:2209.02976 (2022)

    2022 arXiv

  4. [12]

    arXiv preprint arXiv:1806.09055 (2018)

    Liu, H., Simonyan, K., Yang, Y.: Darts: Differentiable architecture search. arXiv preprint arXiv:1806.09055 (2018)

    2018 arXiv

  5. [13]

    In: ECCV (2016)

    Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.Y., Berg, A.C.: Ssd: Single shot multibox detector. In: ECCV (2016)

    2016

  6. [14]

    In: ICPR (2021)

    Ouaknine,A.,Newson,A.,Rebut,J.,Tupin,F.,Pérez,P.:Carradadataset:Camera and automotive radar with range- angle- doppler annotations. In: ICPR (2021)

    2021

  7. [15]

    In: CVPR (2021)

    Qian, K., Zhu, S., Zhang, X., Li, L.E.: Robust multimodal vehicle detection in foggy weather using complementary lidar and radar signals. In: CVPR (2021)

    2021

  8. [16]

    In: CVPR (2016)

    Redmon, J.: You only look once: Unified, real-time object detection. In: CVPR (2016)

    2016

  9. [17]

    In: CVPR (2017)

    Redmon, J., Farhadi, A.: Yolo9000: better, faster, stronger. In: CVPR (2017)

    2017

  10. [18]

    IEEE TPAMI (2017)

    Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object de- tection with region proposal networks. IEEE TPAMI (2017)

    2017

  11. [19]

    arXiv preprint arXiv:2502.12524 (2025)

    Tian, Y., Ye, Q., Doermann, D.: Yolov12: Attention-centric real-time object detec- tors. arXiv preprint arXiv:2502.12524 (2025)

    2025 arXiv

  12. [20]

    arXiv preprint arXiv:2405.14458 (2024)

    Wang, A., Chen, H., Liu, L., Chen, K., Lin, Z., Han, J., Ding, G.: Yolov10: Real- time end-to-end object detection. arXiv preprint arXiv:2405.14458 (2024)

    2024 arXiv

  13. [21]

    In: ECCV (2025)

    Wang, C.Y., Yeh, I.H., Mark Liao, H.Y.: Yolov9: Learning what you want to learn using programmable gradient information. In: ECCV (2025)

    2025

  14. [22]

    In: CRV (2021)

    Zhang, A., Nowruzi, F.E., Laganiere, R.: Raddet: Range-azimuth-doppler based radar object detection for dynamic road users. In: CRV (2021)

    2021

  15. [23]

    In: ICMR (2021)

    Zheng, Z., Yue, X., Keutzer, K., Sangiovanni Vincentelli, A.: Scene-aware learning network for radar object detection. In: ICMR (2021)

    2021

  16. [24]

    In: CVPR (2020)

    Zhou, D., Zhou, X., Zhang, W., Loy, C.C., Yi, S., Zhang, X., Ouyang, W.: Econas: Finding proxies for economical neural architecture search. In: CVPR (2020)

    2020

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.