{"id":"78919be1-7da7-4344-bd72-148583c29a96","arxiv_id":"2509.01280","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A YOLOv8-based detector with a multi-representation heatmap/grayscale adapter branch and one-shot NAS reports new SOTA mAP@50 of 71.9 (RADDet) and 57.1 (CARRADA).","lead":"This paper builds an efficient radar object detector that feeds the same Range-Doppler map into two branches as a heatmap and a grayscale image, and uses neural architecture search to slim the adapter branch. It reports the best published mAP@50 on the RADDet and CARRADA benchmarks (71.9 and 57.1) with fewer than 3M parameters.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RADDet SOTA margin rests on unverified RD-spectrogram comparability; prior-published baseline numbers may not transfer to the authors' self-generated FFT inputs.","rationale":"The paper's contribution is an architecture plus NAS, and the internal evidence is fairly consistent: the Adapter improves across YOLO variants, NAS finds lighter and better subnets, and the fusion ablation shows the proposed fusion as best. The exclusive weak point is the external SOTA claim. The reader identified the same preprocessing comparability issue for RADDet; I agree. It is a verification gap rather than an internal inconsistency: the authors can settle it by releasing preprocessing code and rerunning baselines. Since the reader's verdict is already CONDITIONAL and this concern does not push beyond that, no change in verdict is needed. The CARRADA result's mAP@30 deficit weakens the unqualified wording 'new state-of-the-art,' but it is not an independent fatal issue.","tokens_in":12071,"tokens_out":4914,"duration_ms":60025,"concrete_test":"Run the official DAROD and FCOS+RiCL implementations on the authors' generated RADDet RD spectrograms, using the same train/val split, input representation, and NMS settings; if reproduced mAP@50 remains near the published 46.6/41.0, preprocessing is not the cause, but if it rises materially toward 71.9, the SOTA claim is an artifact. Additionally, check the RADDet dataset documentation to confirm raw ADC availability and compare FFT parameters/statistics with the official RD maps.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central SOTA claim (Section 4.2) depends on comparing YOLOv8n-Adapter-Search against DAROD and FCOS+RiCL numbers taken from prior papers. For RADDet, Section 4.1 says the authors 'extract the raw analog-to-digital-converter data and process it using range and Doppler FFT to generate RD spectrograms.' No FFT length, windowing, normalization, dynamic range, CFAR/quantization, or train/val split details are given, and no preprocessing code is released. The grayscale/heatmap inputs fed to the new model are therefore not shown to be the same object DAROD and RiCL saw. If the custom FFT pipeline yields cleaner or higher-contrast maps, a large part of the 25.3 (vs DAROD) and 30.9 (vs FCOS+RiCL) mAP@50 margins could be a preprocessing artifact. CARRADA is less decisive: official RD spectrograms are used, but the mAP@50 gain over DAROD is only +1.3 and the model is 8.6 points lower at mAP@30. The internal ablations (Tables 4-7) support the Adapter design, so the issue is specifically the cross-paper SOTA comparison, not the architecture.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":12431,"tokens_out":4190,"duration_ms":51244,"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":[{"comment":"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.","section":"§4.1, Table 2"},{"comment":"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.","section":"Table 2, CARRADA row"},{"comment":"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.","section":"§3.4, §4.1"}],"minor_comments":[{"comment":"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.","section":"§4.1"},{"comment":"Typo: 'supassing' should be 'surpassing'.","section":"§4.2"},{"comment":"'ObjectionDetection' should be 'Object Detection'.","section":"Keywords"},{"comment":"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.","section":"Table 1"},{"comment":"No code or preprocessing script is mentioned. Releasing the RD-generation script for RADDet would address the main reproducibility concern and strengthen the paper.","section":"General"},{"comment":"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.","section":"Figure 2 / Table 3"}],"recommendation":"major_revision","confidential_remarks":"The internal architecture work is credible and the ablations are informative. The main risk is that the RADDet SOTA claim is inflated by input-preprocessing differences; if the authors cannot substantiate comparability, the SOTA claim should be removed or substantially qualified. The paper fits the scope of a computer-vision venue, but the current evidence supports an efficiency-architecture contribution more strongly than a new SOTA on RADDet."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague, quick take: this is a workmanlike architecture paper with honestly reported ablations, but the big claim—25 mAP over DAROD on RADDet—needs a re-run or a released preprocessing pipeline before I'd believe it. The internal evidence is decent; the external comparison is the weak link.\n\nWhat's actually new: the combination of heatmap + grayscale RD inputs, the adapter branch with the exchanger (mode 1/2) and primary-auxiliary fusion, plus one-shot NAS on the adapter width/fusion options. None of these parts is new on its own, but the package is new to RD radar detection, and the ablations (Tables 4-7) show consistent gains: adapter gets +2.9 mAP@50-95 on RADDet over YOLOv8n, +1.6 on CARRADA, and the generalization across YOLOv8/9/10 (Table 4) suggests the design is not a fluke. I also credit the search results (Table 3) for reporting params and mAP@50-95; the Pareto plots in Fig 5 are standard but fine.\n\nThe problem is the SOTA table. Section 4.1 states they 'extract the raw ADC data and process it using range and Doppler FFT' for RADDet, and the paper never reports FFT length, windowing, normalization, dynamic range, or train/val split details. DAROD and RiCL numbers in Table 2 come from earlier papers. If those papers used different RD maps—especially cleaner or higher contrast—the 25.3-point gap could be mostly preprocessing. That's not a minor concern; it's the central claim. CARRADA is less central: official spectrograms are used, and the mAP@50 gain over DAROD is only +1.3, with an 8.6-point deficit at mAP@30.\n\nThere are smaller soft spots: the top-5 subnet selection criterion isn't specified (val or test?), no error bars or code, and the NMS IoU of 0.1 seems unusual but it's their stated choice. The circularity concern is not real: the NAS selection is best-of-five full training runs; I'd call that a mild selection effect, not evidence of fitting.\n\nWho this is for: researchers working on RD-radar object detection who want an efficient dual-branch architecture and a benchmark to chase. It deserves a proper peer-review round—not because the SOTA numbers are credible yet, but because the architecture and internal ablations are actually informative. The authors should be asked to release the RD generation code, re-run baselines on their own maps, and report selection criteria and variance. I'd accept it for peer review and let the revision process sort the preprocessing question.\n\nMy recommendation: engage with it, but take the RADDet SOTA with a grain of salt until the preprocessing is public.","headline":"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.","tokens_in":12871,"tokens_out":2468,"would_cite":true,"duration_ms":29877,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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","keywords":["Range-Doppler radar","object detection","multi-representation","adapter branch","neural architecture search","YOLO","heatmap","grayscale"],"falsifier":"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.","tokens_in":12008,"feed_emoji":"📡","tokens_out":8243,"duration_ms":85410,"temperature":0.7,"pith_summary":"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.","feed_headline":"Heatmap + grayscale radar combo tops RD detection at 2.65M params","feed_subtitle":"The searched two-branch detector reports mAP@50 of 71.9 on RADDet and 57.1 on CARRADA.","key_machinery":"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","core_discovery":"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","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"YOLOv8 is the base detector whose backbone, neck, and head host the searchable adapter; removing it removes the whole pipeline.","marker":"[9]"},{"why":"DAROD is the main prior RD-map detector and baseline; it also supplies the evaluation protocol (NMS IoU threshold 0.1) the paper follows.","marker":"[2]"},{"why":"RADDet supplies the benchmark and the raw ADC data from which the paper generates RD spectrograms; it also appears as a high-parameter baseline.","marker":"[22]"},{"why":"CARRADA supplies the second benchmark and its official RD spectrograms, used to test generalization beyond the authors' own preprocessing.","marker":"[14]"},{"why":"FCOS+RiCL is the semi-supervised contrastive baseline the paper compares against and whose complex pretraining pipeline the method is meant to avoid.","marker":"[3]"},{"why":"Coordinate Attention is the lightweight attention operation used inside every Exchanger Module to gate the fused features.","marker":"[7]"},{"why":"The single-path one-shot NAS method supplies the uniform-sampling weight-sharing supernet training procedure that lets all subnets inherit one set of weights.","marker":"[5]"},{"why":"The evolutionary NAS methodology (population, crossover, mutation, constraints) is used to search the supernet for the best subnet under parameter constraints.","marker":"[24]"}],"fun_headline_variants":["Radar maps: two views beat one for efficient detection","Searched adapter fuses heatmap + grayscale for radar SOTA","Dual radar views with NAS top RD detection at 2.65M","Efficient radar detector: heatmap + grayscale with auto-search","Multi-view radar adapter: small model, SOTA results"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Radar maps: two views beat one for efficient detection","Searched adapter fuses heatmap + grayscale for radar SOTA","Dual radar views with NAS top RD detection at 2.65M","Efficient radar detector: heatmap + grayscale with auto-search","Multi-view radar adapter: small model, SOTA results"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000152,"raw_usage":{"total_tokens":1033,"prompt_tokens":729,"completion_tokens":304,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":213}},"tokens_in":473,"tokens_out":304,"duration_ms":3871,"temperature":1.0,"reasoning_tokens":213,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:41:27.590525+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"In: IEEE IV (2022)","cited_arxiv_id":null,"evidence_quote":"DAROD is the main prior RD-map detector and baseline; it also supplies the evaluation protocol (NMS IoU threshold 0.1) the paper follows."},{"cited_title":"In: CRV (2021)","cited_arxiv_id":null,"evidence_quote":"RADDet supplies the benchmark and the raw ADC data from which the paper generates RD spectrograms; it also appears as a high-parameter baseline."},{"cited_title":"In: ICPR (2021)","cited_arxiv_id":null,"evidence_quote":"CARRADA supplies the second benchmark and its official RD spectrograms, used to test generalization beyond the authors' own preprocessing."},{"cited_title":"Leveraging Self-Supervised Instance Contrastive Learning for Radar Object Detection","cited_arxiv_id":"2402.08427","evidence_quote":"FCOS+RiCL is the semi-supervised contrastive baseline the paper compares against and whose complex pretraining pipeline the method is meant to avoid."},{"cited_title":"In: CVPR (2021)","cited_arxiv_id":null,"evidence_quote":"Coordinate Attention is the lightweight attention operation used inside every Exchanger Module to gate the fused features."},{"cited_title":"In: ECCV (2020)","cited_arxiv_id":null,"evidence_quote":"The single-path one-shot NAS method supplies the uniform-sampling weight-sharing supernet training procedure that lets all subnets inherit one set of weights."},{"cited_title":"In: CVPR (2020)","cited_arxiv_id":null,"evidence_quote":"The evolutionary NAS methodology (population, crossover, mutation, constraints) is used to search the supernet for the best subnet under parameter constraints."}],"review_version":1}