{"id":"e4852329-c585-4dec-85e1-a7ec4c350de8","arxiv_id":"2502.01940","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A Bartlett-inspired spectral encoding of radar and camera images lets a ResNet generate denser radar depth maps, with reported improvements in Chamfer distance and absolute error on RaDelft.","lead":"The paper introduces a spectral encoding method that lets camera images help train a radar depth map generator, producing denser radar depth maps for autonomous vehicles. If it holds up, it points toward cheaper self-driving perception that does not need lidar.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No inverse spectrum-to-depth transform is defined; Table III metrics are unverifiable, undercutting the central SOTA claim.","rationale":"The reader's weakest_assumption correctly identifies the missing inverse transform as the most load-bearing issue. This is not a minor omission: the paper's only quantitative evidence for the SOTA improvement depends on converting spectrum-domain outputs back into depth maps and point clouds. Without that inverse, the reported MAE, REL, UCD, and BCD values in Table III cannot be checked, and the central claim is effectively unverifiable from the manuscript alone. I considered whether the abstract's metric swap (24.24% vs 52.59% for UCD vs MAE) might be more severe, but that is a consistency error that could be fixed by editing; the missing inverse is a structural gap in the method. I also note that Eq. 8 appears dimensionally inconsistent, which further weakens the description of the forward transform itself, but the inverse problem remains the clearest blocker. There is no independent support (e.g., machine-checked proofs or reproducible code in the manuscript) that mitigates this gap. Therefore, I agree with the reader's rejection and do not see a need to adjust the verdict.","tokens_in":9948,"tokens_out":3833,"duration_ms":36193,"concrete_test":"Obtain the authors' GitHub repository and identify the exact function that converts a ResNet output spectrum image into the depth maps used for MAE/REL and the point clouds used for UCD/BCD. Re-run inference with that function on a held-out frame and recompute Table III. If no inverse function exists, attempt to recover depth via any plausible inversion (e.g., projecting onto the Bartlett basis or using the raw power as a pseudo-depth); if the reported metrics cannot be reproduced, the central claim is unverifiable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the proposed spectrum-learning pipeline reduces MAE, REL, UCD, and BCD versus SOTA on RaDelft. For that claim to be meaningful, the ResNet output, which lives in the spatial-spectrum domain according to Eq. 11, must be mapped back to a depth map or point cloud against which MAE/REL/UCD/BCD are computed. The paper never specifies this inverse mapping. Section V.B merely states that \"spectrum images are first transformed into 3D point clouds\" for UCD/BCD, but no such transform appears in Section IV, V, or any appendix. Even the forward transform Eq. 8 is dimensionally ambiguous: the right-hand side sums over all n and k, yielding a scalar, while P is declared as an N-by-K matrix. Without an explicit inverse, one cannot reproduce Table III, verify the claimed 24.24% MAE and 52.59% UCD improvements, or even determine whether the \"sharper depth maps\" in Fig. 5 are the raw ResNet output or a post-processed product. This gap is load-bearing because the entire quantitative evaluation rests on it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a camera-radar fusion pipeline for generating denser 4D radar depth maps. RGB images are passed through a semantic segmentation network, and both the radar depth map and the segmentation map are transformed by a Bartlett-inspired nonlinear spatial-spectrum encoding. A ResNet is trained to map the radar spectrum to the element-wise product of the radar and camera spectra, with L2 loss. After offline training, the authors claim the radar-only model can produce sharper depth maps that outperform the SOTA detector on RaDelft, reporting reductions in MAE, REL, UCD, and BCD. The paper also reports increased Pearson correlation and mutual information between the encoded modalities and presents an ablation study over the number of spectral segments M.","tokens_in":10220,"tokens_out":6198,"duration_ms":63888,"significance":"If the central claims held, the idea of learning in a common spectral subspace shared by radar and camera would be an interesting step toward low-cost, lidar-free depth perception for autonomous vehicles. The paper uses a public dataset, compares against a recent SOTA detector, and includes ablations over the key parameter M, which are positive features. However, the significance cannot currently be assessed because the evaluation pipeline is incomplete: no inverse transform from the spectral domain to depth maps or point clouds is specified, no train/test split is described, and hyperparameters appear to be selected on the same frames used for final evaluation. These are load-bearing gaps, not presentation issues, because Table III and the claimed SOTA improvements rest entirely on the missing components.","major_comments":[{"comment":"The inverse transform from the learned spatial spectrum to depth maps or point clouds is never defined. The network is trained to predict P_radar ∘ P_cam in the spectral domain, but MAE, REL, UCD, and BCD in Table III are computed on depth maps and point clouds. The sentence in Section V.B that \"spectrum images are first transformed into 3D point clouds\" is the only description, and no such transform appears in Section IV or anywhere else. Without an explicit inverse of F(·) or H(·), the quantitative results in Table III cannot be reproduced, and the claimed 24.24%, 18.52%, 52.59%, and 10.41% improvements are unverifiable from the manuscript.","section":"Section V.B and Eq. (11)"},{"comment":"The forward transform is not well-defined as written. Equation (8) states P(n,k) = Σ_{n=0}^{N-1} Σ_{k=0}^{K-1} |Y(ϕ_n,θ_k) ∘ I|, but the right-hand side sums over n and k, so it is a scalar independent of the left-hand side indices, making P(n,k) inconsistent with the declared N×K matrix P. Additionally, the dimensions of I and of Y are not specified, so it is unclear how an input image is encoded into the spectrum. This ambiguity undermines the definition of the core preprocessing step used throughout the paper.","section":"Section IV.A, Eq. (8)"},{"comment":"No train/test split is described. The paper states that scenes 2 and 6 of RaDelft include 5680 frames and that these are used for the experiments, but it does not state how many frames are used for training, validation, and testing. If the metrics in Table III are computed on the same frames used for training, they measure fitting rather than generalization. The paper must specify the split and report metrics on a held-out test set for the claims to be meaningful.","section":"Section V (experimental setup)"},{"comment":"The hyperparameters M_radar and M_cam appear to be selected on the same frames later used for evaluation. Figure 7 is used to choose M_radar = 20 and M_cam = 200 by maximizing Pearson correlation and mutual information, and Table III is then reported on the same dataset. This is a form of test-data leakage that can inflate the reported improvements. The paper should select hyperparameters on a validation split and clearly separate that selection from the final evaluation.","section":"Section VI.A and Table III"},{"comment":"The training objective contains a potentially trivial component: the target P_radar ∘ P_cam includes P_radar as a multiplicative factor, and the network input is P_radar itself. A network can reduce part of the L2 loss by approximating the identity on P_radar, regardless of the camera information. The paper does not isolate how much of the reported depth-map improvement comes from the camera term versus this self-predictive component. An ablation that trains with P_radar as both input and target, or with the camera term removed, would be needed to support the claim that cross-modal learning is responsible for the gains.","section":"Section IV.D, Eq. (11)"}],"minor_comments":[{"comment":"The headline improvement numbers are inconsistent. The abstract at the top of the paper says \"outperforms the state-of-the-art (SOTA) by 27.95% in terms of Unidirectional Chamfer Distance (UCD)\", while the full-text abstract and Introduction report different pairings; Table III lists 52.59% UCD and 24.24% MAE. These numbers should be reconciled.","section":"Abstract and Introduction"},{"comment":"The paper says experiments were performed for M ∈ {10, 20, 50, 70, 200}, but Section VI.A says the ablation uses M ∈ {10, 12, 14, ..., 200}. Please clarify which values were actually used and how Figure 7 was generated.","section":"Section V.A"},{"comment":"The spectrum-representation ablation in Figure 8 is qualitative only. A quantitative comparison, using the same metrics as Table III, would be needed to support the claim that the spectrum step is essential.","section":"Section VI.C"},{"comment":"Training details are missing: the paper does not specify the optimizer, learning rate schedule, batch size, loss weighting, or hardware. The GitHub link is useful, but the manuscript should contain enough detail for independent reproduction.","section":"Section V.B"},{"comment":"There are numerous typographical and notation inconsistencies, including \"V oD\" instead of \"VoD\", \"A Vs\" spacing, \"SoTA\" vs \"SOTA\", and the undefined symbol \"F(.)\" in Figure 1. A careful editing pass is needed.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The core problem is that the quantitative evaluation is not reproducible from the manuscript: the spectral-to-depth inverse is missing, the forward transform in Eq. (8) is dimensionally inconsistent, and the experimental protocol lacks a train/test split. These are not cosmetic issues; they affect the validity of every reported improvement. If the authors can supply the missing inverse transform, a proper experimental protocol, and an ablation that separates the camera contribution from the self-predictive part of the loss, a substantially revised manuscript might be reconsidered, but the current submission does not support its central claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the non-linear phase-progression basis and the spectrum-domain training target are genuinely new relative to the cited radar-camera fusion work, and the cross-modal correlation/mutual-information gains are plausible. But the headline SOTA numbers cannot be checked, because the inverse spectral-to-depth transform is never given. That is load-bearing, not a footnote.\n\nWhat the paper does well: the pixel positional encoding via a non-linear Bartlett-style basis is a real contribution, and using the element-wise product P_radar * P_cam as a training target is not something I have seen in the prior work they cite. The ablation study is a good-faith attempt, and the qualitative results in Fig. 5 do suggest sharper depth maps. The framing of the deployment scheme — camera supervises radar offline, radar runs solo online — is sensible and worth exploring.\n\nWhere it falls apart: Section V.B says \"spectrum images are first transformed into 3D point clouds\" for UCD/BCD, but no such inverse transform appears in Section IV, V, or the appendix. Eq. 8 is dimensionally off: the RHS sums over n and k, yielding a scalar, while P is declared an N-by-K matrix. Without the inverse mapping, a reader cannot reproduce Table III, verify the claimed 24.24% MAE / 52.59% UCD reductions, or even tell whether Fig. 5 shows raw ResNet output or a post-processed product. The stress-test note is on point.\n\nThe abstract also disagrees with the main text (it says 27.95% UCD and swaps the MAE/UCD percentages), which does not inspire confidence in the reporting. On top of that, no train/test split is described; all 5680 frames from scenes 2 and 6 are used, and the M_radar/M_cam hyperparameters are selected using the same evaluation frames via the Fig. 7 analysis. The paper also does not compare against the radar-camera fusion methods [12]–[16] that it lists in related work, which makes the \"SOTA\" claim weaker than it should be.\n\nBottom line: the core ideas deserve a serious referee round, but this version is not publishable as is. Major revision is needed: specify the inverse transform, fix the equation, report a real train/test split, and add at least one strong baseline from the cited fusion literature. I would not desk-reject it, but I would not accept it either.","headline":"Clever spectral encoding idea, but the main quantitative claim is unverifiable as written: no inverse transform from the spectrum domain back to depth maps is specified.","tokens_in":10682,"tokens_out":2802,"would_cite":false,"duration_ms":27220,"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 shows that a camera used only during offline training can teach a 4D radar to produce sharper depth maps, beating the state-of-the-art detector by up to 52.59% in unidirectional chamfer distance.","keywords":["autonomous vehicles","4D radar","depth map generation","spatial spectrum estimation","camera-radar fusion","semantic segmentation","positional encoding","depth completion"],"falsifier":"Run the trained pipeline on a held-out scene of the same 4D radar dataset and compare the output depth maps to lidar ground truth; if the reported unidirectional chamfer distance (UCD) and mean absolute error (MAE) improvements over the state-of-the-art detector do not appear on that scene, the claimed generalization is not established.","tokens_in":9777,"feed_emoji":"🚗","tokens_out":15019,"duration_ms":129747,"temperature":0.7,"pith_summary":"This paper tries to establish that a low-cost camera can be used offline to train a 4D radar sensor to produce dense, sharp depth maps for autonomous vehicles. The method transforms the radar depth map and a semantic segmentation of the camera image into a shared 'spatial spectrum' subspace, using a nonlinear pixel positional encoding inspired by periodogram-averaging spectral estimation, then trains a residual network to predict the elementwise product of the two spectra from the radar spectrum alone. In the paper's experiments on two scenes of a public 4D radar dataset, the resulting depth maps reduce mean absolute error by 24.24%, relative error by 18.52%, unidirectional chamfer distance (a point-cloud distance metric) by 52.59%, and bidirectional chamfer distance by 10.41% compared with the state-of-the-art data-driven radar detector. The encoding also raises linear correlation and mutual information between the two modalities by factors of 3.88 and 76.69. If the claim holds, lidar could be removed from the sensor stack without giving up radar's all-weather robustness.","feed_headline":"Camera spectra cut radar depth-map error by 52.6%","feed_subtitle":"After offline training, the radar runs camera-free, pointing toward cheaper, lidar-free autonomous driving.","key_machinery":"The load-bearing object is the nonlinear spatial spectrum transform. Starting from $M$ segments with phase progression $x(m,\\phi) = e^{-j\\pi m\\sin(\\phi)}$ and $x(m,\\theta) = e^{-j\\pi m\\sin(\\theta)}$, the algorithm forms periodogram estimates $y(\\phi_n)$ and $y(\\theta_k)$, combines them into a joint 2D periodogram $Y(\\phi_n,\\theta_k)$, and accumulates $P(n,k) = \\sum_{n,k} |Y(\\phi_n,\\theta_k) \\circ I|$ over all pixels of the input image. The resolution of the resulting spectrum grows with the number of segments $M$, and spectral leakage introduces high-frequency components that help complete sparse depth and fit sharp camera contours. This transform is what lets the training network compare radar and camera data in a common subspace, rather than trying to map between raw depth and RGB pixel spaces directly.","core_discovery":"The central claim is that radar and camera images can be brought into one learnable subspace, the spatial spectrum, by estimating each image's 2D power spectrum over a common set of nonlinear spatial-frequency basis functions. For the radar, the preprocessing computes $P_{radar} = F(I_{radar}, M_{radar})$; for the camera, the RGB frame is first segmented semantically and then encoded as $P_{cam} = F(Seg, M_{cam})$, with $M_{cam} > M_{radar}$ so the camera side retains higher resolution. Training uses the objective $P_{radar} \\circ P_{cam} = \\mathrm{ResNet}(P_{radar})$ under L2 loss, so the network learns to reconstruct the elementwise product of radar and camera spectra from the radar input alone. The paper reports that this yields sharper object contours and lowers MAE, REL, UCD, and BCD relative to the state-of-the-art data-driven radar detector, and that after offline training the camera is no longer needed at runtime.","pith_inferences":["If the spectrum transform is modality-agnostic, the same encoding could sharpen lidar or sonar depth maps using camera supervision offline, extending the approach beyond automotive radar.","Because the camera is removed at runtime, the method's practical value depends on weather robustness: a natural test is measuring depth-map accuracy in rain, fog, or darkness, where camera-based supervision would be degraded or absent.","The paper's argument that the segmentation benchmark matters more than the segmentation model implies that transferring to a new city or object set requires re-training the segmentation stage on a matching urban benchmark.","Reproducing the quantitative claims requires the authors to state the missing inverse spectrum transform; a follow-up with a differentiable, explicitly invertible encoding would make the whole pipeline verifiable."],"forward_implications":["After offline training, the 4D radar depth generator runs without the camera, so a production system could keep radar's all-weather robustness while shedding lidar cost.","Denser, contour-accurate radar depth maps should improve downstream perception, tracking, and rendering modules that consume point clouds.","The large gains in linear correlation and mutual information after encoding suggest the spatial spectrum is a promising common representation for other sensor-modality pairs, not just radar and camera.","The ablation results tie performance to the resolution parameters $M_{radar} \\approx 20$ and $M_{cam} \\ge 180$, so each new sensor pairing would need its own resolution search."],"supporting_citations":[{"why":"Provides the data-driven 4D radar detector whose sparse depth maps are the input to the proposed pipeline and serves as the state-of-the-art baseline.","marker":"[6]"},{"why":"The periodogram-averaging spectral estimation method that motivates the paper's spatial spectrum transform.","marker":"[8]"},{"why":"The semantic segmentation network that converts each RGB frame into class-level segmentations before the camera-side encoding.","marker":"[22]"},{"why":"The residual network architecture used as the backbone for segmentation and as the depth-map generator in training.","marker":"[7]"},{"why":"The urban-scene semantic segmentation benchmark whose class coverage determines which objects the camera spectrum can encode.","marker":"[23]"}],"fun_headline_variants":["Radar sharpens via camera spectra, cutting depth error by 28%","Radar learns from camera spectrum, drops depth-map error 28%","Spectrum-based fusion yields denser radar depth maps for AVs","Camera teaches radar to sharpen depth maps—no camera at runtime","One shared spectral space lets radar match camera depth detail"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline rests on the unstated assumption that improving accuracy in the transformed spectrum space automatically improves the final depth maps, because the paper does not describe the inverse transform that turns the network's spectrum output back into depth coordinates.","fun_headline_variants_meta":{"raw":{"variants":["Radar sharpens via camera spectra, cutting depth error by 28%","Radar learns from camera spectrum, drops depth-map error 28%","Spectrum-based fusion yields denser radar depth maps for AVs","Camera teaches radar to sharpen depth maps—no camera at runtime","One shared spectral space lets radar match camera depth detail"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000634,"raw_usage":{"total_tokens":2923,"prompt_tokens":939,"completion_tokens":1984,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":1894}},"tokens_in":555,"tokens_out":1984,"duration_ms":15996,"temperature":1.0,"reasoning_tokens":1894,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T13:55:18.312815+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained pipeline on a held-out scene of the same 4D radar dataset and compare the output depth maps to lidar ground truth; if the reported unidirectional chamfer distance (UCD) and mean absolute error (MAE) improvements over the state-of-the-art detector do not appear on that scene, the claimed generalization is not established.","supporting_citations":[{"cited_title":"See Further Than CFAR: a Data-Driven Radar Detector Trained by Lidar","cited_arxiv_id":"2402.12970","evidence_quote":"Provides the data-driven 4D radar detector whose sparse depth maps are the input to the proposed pipeline and serves as the state-of-the-art baseline."},{"cited_title":"Smoothing periodograms from time-series with contin- uous spectra,","cited_arxiv_id":null,"evidence_quote":"The periodogram-averaging spectral estimation method that motivates the paper's spatial spectrum transform."},{"cited_title":"Semantic segmentation with extended deeplabv3 architecture,","cited_arxiv_id":null,"evidence_quote":"The semantic segmentation network that converts each RGB frame into class-level segmentations before the camera-side encoding."},{"cited_title":"The cityscapes dataset for semantic urban scene understanding,","cited_arxiv_id":null,"evidence_quote":"The urban-scene semantic segmentation benchmark whose class coverage determines which objects the camera spectrum can encode."}],"review_version":1}