{"id":"2408a52d-91ac-4410-954f-f405f8687b11","arxiv_id":"2506.15560","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"A radar-camera depth estimation framework that recalibrates sparse radar points and aligns a frozen monocular depth model using sparse LiDAR labels, claiming state-of-the-art accuracy with roughly 1% supervision density.","lead":"RaCalNet is a new system that estimates dense metric depth from 4D radar and a single camera using only about 1% of the LiDAR supervision required by prior methods. It screens and shifts radar points before aligning a frozen monocular depth model to them, and the authors report matching or beating dense-supervised baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported RMSE reductions contradict Tables II and III, and the public SOTA claim depends on an unreleased ground-truth correction, so the central quantitative evidence is not reproducible.","rationale":"I read the paper as proposing RaCalNet, a sparse-supervised radar depth estimator that claims to match or beat dense-supervised methods. For that claim to hold, the evaluations must be internally consistent and reproducible. They are not: the percentage reductions quoted in the text do not match the tables, and the public-dataset result depends on an unreleased ground-truth correction. The affine assumption in Eqs. (11)/(14) is a plausible methodological risk, but it is secondary to the empirical contradictions. Even if the affine assumption were perfect, the numbers as reported do not support the headline. I also note the paper retrains the baseline on the corrected dataset and drops the other baselines (DORN, Singh) from the corrected table, so 'state-of-the-art' is asserted without a full comparison. The real-world dataset is proprietary, so the 34.89% cannot be audited. These issues warrant the reader's REJECT verdict; I do not see a reason to change it. The one step that would materially help is releasing the corrected ground truth and the calibration code, and recomputing the claimed percentages from the tables.","tokens_in":14007,"tokens_out":6566,"duration_ms":64819,"concrete_test":"Independently recompute all RMSE reduction percentages from Tables II and III using the formula (RMSE_baseline − RMSE_ours)/RMSE_baseline and compare them with the percentages stated in Sections IV-C and IV-D; the current values do not match, which already indicates an internal inconsistency. Additionally, request the authors to release the extrinsic-calibration code and corrected ZJU ground-truth maps, and to rerun RaCalNet and Radarcam on the original uncorrected ZJU test split; the SOTA claim is only verifiable if it holds on the unaltered public benchmark.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-C claims RMSE reductions of 35.08%, 29.12%, and 27.27% at 50/70/80 m on the corrected ZJU evaluation, but recomputing from Table II gives (1 - 2137.652/3408.806) ≈ 37.3%, (1 - 2358.958/3719.426) ≈ 36.6%, and (1 - 2479.008/3831.740) ≈ 35.3%. Section IV-D claims 46.24%, 43.42%, and 40.06% on the real-world dataset, but Table III yields ≈ 38.8%, 36.4%, and 34.9. The abstract's 34.89% matches the 80 m row of Table III, not the text. These discrepancies make the headline quantitative claim internally inconsistent. Moreover, the public-benchmark advantage appears only after an under-specified extrinsic correction of the ZJU ground truth; the original evaluation (Table I) shows RaCalNet with lower RMSE/SqRel but worse MAE/AbsRel/δ1 than Radarcam, not clearly SOTA. The correction is not described with enough detail (optimization objective, initial values, whether LiDAR or radar extrinsics were adjusted) and neither the corrected ground truth nor code is released. Since the proprietary real-world dataset is unavailable, no public evidence remains to verify the claimed advantage.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"RaCalNet proposes a radar-camera depth estimation framework that avoids dense LiDAR supervision. A Radar Recalibration module, trained with sparse single-frame LiDAR, predicts per-point confidence scores and pixel displacements for raw 4D radar projections, producing refined sparse depth anchors. A Metric Depth Optimization module then aligns frozen monocular inverse-depth predictions to these anchors via global affine least squares, K-means cluster-wise affine refinement, and edge-aware smoothing. The paper reports state-of-the-art results on the ZJU-4DRadarCam dataset after correcting its ground-truth extrinsics, and large RMSE reductions on a proprietary real-world dataset, alongside ablations and a downstream 3D reconstruction demonstration.","tokens_in":14334,"tokens_out":4084,"duration_ms":47077,"significance":"If substantiated, the sparse-supervision result would be practically significant: reducing supervision density to roughly 1% of dense LiDAR methods while retaining metric accuracy could lower data collection cost for radar-camera depth estimation. The modular design, with sparse LiDAR supervising radar recalibration rather than depth directly, is a reasonable and interesting idea, and the ablation study in Table IV gives some evidence that both screening and displacement refinement contribute. However, the reported quantitative evidence is internally inconsistent, and the public-dataset superiority depends on an under-specified, unreleased ground-truth correction, which prevents verification of the central claims.","major_comments":[{"comment":"The reported RMSE reductions are not consistent with the tables. In Section IV-C the text claims reductions of 35.08%, 29.12%, and 27.27% at 50m, 70m, and 80m, but Table II yields (1 - 2137.652/3408.806) = 37.3%, (1 - 2358.958/3719.426) = 36.6%, and (1 - 2479.008/3831.740) = 35.3%. In Section IV-D the text claims 46.24%, 43.42%, and 40.06%, but Table III gives 38.8% (50m), 36.4% (70m), and 34.9% (80m). The abstract's 34.89% figure matches only the 80m row of Table III, not the text. Because these percentages are the headline quantitative contribution, this discrepancy undermines the paper's central claim.","section":"Section IV-C, IV-D, Abstract"},{"comment":"The claimed state-of-the-art result on ZJU depends entirely on a ground-truth correction that is not adequately specified or released. On the original evaluation (Table I), RaCalNet has lower RMSE and SqRel than Radarcam but worse MAE, AbsRel, and δ1; it is not clearly superior. The corrected evaluation (Table II) appears only after \"optimizing the extrinsic parameters and reprojecting the LiDAR data,\" with no objective function, no initialization, no details of which extrinsics were adjusted, and no release of the corrected ground truth. Since the paper's public-benchmark advantage is not reproducible without this correction, the evidence cannot be independently checked.","section":"Section IV-C, Table I, Fig. 4"},{"comment":"The Metric Depth Optimization module assumes that the frozen monocular network's inverse depth is related to true metric inverse depth by an affine transformation, both globally in Eq. (11) and per cluster in Eq. (14). This is a load-bearing assumption: if the monocular model has non-affine inverse-depth errors, the least-squares fits will be biased, and the bias will propagate into the final metric depth. The paper does not test this assumption separately, for example by comparing fitted residuals on held-out radar points against a more flexible calibration model, so the robustness of the core optimization step is unverified.","section":"Section III-B, Eqs. (11) and (14)"}],"minor_comments":[{"comment":"The text says \"Fig. 5 presents visual comparisons organized in five rows,\" but the figure caption and displayed content show three rows: input images, depth estimations, and error maps.","section":"Section IV-C, Fig. 5"},{"comment":"Reference [34] contains a typo: \"PProceedings of the IEEE Conference\" should be \"Proceedings of the IEEE Conference.\"","section":"References"},{"comment":"The baseline name is inconsistently spelled as both \"Radarcam\" and \"RadarCam\"; the paper should use one consistent form.","section":"Throughout"},{"comment":"The cluster-count heuristic uses the symbol Nvalid, but Nvalid is not defined in Algorithm 1 or the surrounding text; it should be defined explicitly (the reader can infer it is the number of valid radar points, but this should be stated).","section":"Algorithm 1"},{"comment":"The phrase \"the strongest-performing baseline\" is ambiguous in the corrected-dataset comparison, and the name \"DPT+Var+RC-Net\" for Radarcam is introduced without definition or reference to a specific configuration.","section":"Section IV-C"}],"recommendation":"reject","confidential_remarks":"The major issue is not the idea but the evidence: the headline percentages contradict the tables, and the public-benchmark advantage relies on an unreleased corrected ground truth. If the authors release the corrected data, the correction procedure, and the code, and if the corrected numbers are verified, a resubmission could be considered. As it stands, the central quantitative claims are neither internally consistent nor independently verifiable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know first: this paper has a genuinely interesting architecture and a significant claim—radar recalibration can replace dense LiDAR supervision with ~1% sparse labels. The Radar Recalibration module (confidence screening plus learned pixel displacement via cross-modal attention) and the multi-stage metric alignment on a frozen monocular network are both sensible and, as far as I can tell, not present in the cited prior work. The ablation on the proprietary dataset shows the recalibration module matters: removing it roughly doubles MAE. That is real evidence the module does something.\n\nThe soft spots are not trivial, though. The stress-test note is right: the percentages in Section IV-C and IV-D do not match the tables. Recomputing from Table II gives 37.3%, 36.6%, and 35.3% reductions at 50/70/80 m, not the 35.08%, 29.12%, and 27.27% in the text. Table III similarly yields ~38.8%, ~36.4%, and ~34.9%, not 46.24%, 43.42%, and 40.06%. The abstract's 34.89% matches the 80 m row of Table III, not the text. These are not stylistic discrepancies; the central quantitative claim is internally inconsistent.\n\nSecond, the public-benchmark advantage appears only after an under-specified extrinsic correction of the ZJU ground truth. The original Table I shows RaCalNet beating Radarcam on RMSE and SqRel but losing on MAE, AbsRel, and δ1—not clearly SOTA. The correction procedure (objective, initialization, which extrinsics were adjusted) is not described, and neither the corrected ground truth nor code is released. Combined with the proprietary real-world dataset, there is currently no publicly verifiable evidence for the headline claim.\n\nThe metric optimization also leans on an affine-in-inverse-depth assumption (global and per-cluster) that is plausible but never isolated. That is more a limitation to state than a fatal flaw.\n\nWho is this for? Researchers working on radar-camera depth fusion. It deserves a serious referee because the direction is promising and the method is not obviously wrong—but the current submission cannot be accepted without the numbers fixed, the correction documented, and at least the corrected ground truth or code released. I would not cite the quantitative results in their present form, but I would read the revised version carefully.\n\nRecommendation: send it to peer review, but expect heavy revision. The kernel of the idea is worth the referee time.","headline":"The core idea—sparse-supervised radar recalibration plus monocular metric alignment—is worth taking seriously, but the paper's headline numbers don't survive arithmetic, and the only clean public-benchmark win rests on an undisclosed, unreleased ground-truth correction.","tokens_in":14909,"tokens_out":1341,"would_cite":false,"duration_ms":16205,"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":"A two-module network turns sparse, noisy radar points into metric anchors, letting sparse LiDAR supervision replace dense LiDAR supervision in radar-camera depth estimation.","keywords":["4D radar","dense depth estimation","sparse LiDAR supervision","radar-camera fusion","metric depth alignment","cross-modal attention","affine inverse-depth calibration","monocular depth"],"falsifier":"Render or capture a test scene with wide depth variation concentrated in a single K-means cluster (e.g., a long corridor seen at an angle), run the metric depth optimization with and without per-cluster affine fitting, and measure whether the cluster-wise residual error follows the affine model; if residuals are systematically biased by depth, the affine assumption fails.","tokens_in":13759,"feed_emoji":"📡","tokens_out":7723,"duration_ms":76933,"temperature":0.7,"pith_summary":"RaCalNet claims that dense LiDAR supervision—the standard fuel for radar-camera depth estimation—can be replaced by sparse LiDAR at about 1% of the point density, without losing metric accuracy. The paper's route is two-stage: first a cross-modal attention network screens raw radar points for reliability and refines their pixel projections, trained with single-frame sparse LiDAR; then a metric depth optimizer solves for the affine map between a frozen monocular network's inverse depth and the true inverse depth, globally and per cluster, turning the refined radar points into metric anchors. On the ZJU-4DRadarCam benchmark the paper reports state-of-the-art RMSE once the dataset's misaligned ground truth is corrected, and on a proprietary real-world platform it reports RMSE reductions of 40–46% against the dense-supervised Radarcam baseline. If true, this cuts the cost and complexity of radar depth training and avoids the geometric artifacts that come from accumulating and interpolating many LiDAR frames.","feed_headline":"Radar depth from 1% LiDAR supervision rivals dense labels","feed_subtitle":"Recalibrated radar points rescale monocular depth using ~1% of the labels dense methods need.","key_machinery":"The load-bearing object is the affine relationship in inverse-depth space between the frozen monocular network's prediction $\\hat{d}_m$ and the metric depth $d_{\\mathrm{radar}}$ of a refined radar anchor: $\\frac{1}{d_{\\mathrm{radar}}} = \\alpha \\hat{d}_m + \\beta$. The Metric Depth Optimization module solves this globally by regularized least squares, then refits $(\\alpha_c, \\beta_c)$ independently inside each K-means cluster of pixels (clustered by normalized image coordinates and log depth), and finally applies edge-aware Gaussian smoothing outside depth-discontinuity masks. This alignment machinery does the actual work of converting relative monocular depth into metric depth. The Radar Recalibration module is the other half: it supplies the anchors by learning per-point confidence and pixel displacement with cross-modal self/cross-attention between radar features and image patches, supervised by single-frame sparse LiDAR.","core_discovery":"On its own terms, the paper's central discovery is that noise and misprojection in 4D radar points are learnable, and that once learned, the cleaned radar points are sufficient to put metric scale on an otherwise scale-ambiguous monocular depth network. The Radar Recalibration module outputs, for each projected radar point, a confidence score and a 2D displacement, trained against sparse LiDAR via binary cross-entropy and smooth L1 losses; points below a confidence threshold are discarded. The Metric Depth Optimization module then takes the surviving points and solves $1/d_{\\mathrm{radar}} = \\alpha \\hat{d}_m + \\beta$ by regularized least squares, filters unreliable monocular predictions by a searched inverse-depth threshold, refits the affine map per K-means cluster, and smooths non-edge regions. With this pipeline the paper reports RMSE reductions of 35.08%, 29.12%, and 27.27% over the strongest baseline on the corrected ZJU-4DRadarCam test split at 50m/70m/80m, and reductions of 46.24%, 43.42%, and 40.06% on its real-world campus dataset, all while using roughly 1% of the LiDAR supervision density of the dense-supervised comparison.","pith_inferences":["The global-plus-cluster affine alignment is a generic calibration recipe: any scale-ambiguous depth network could be metrically grounded by the same two-step least-squares routine given sparse metric anchors, not just radar ones (e.g., sparse LiDAR, ToF, or visual-inertial points).","The cluster-wise affine refinement effectively fits a piecewise-linear model to the monocular network's error surface; performance could degrade in scenes with very few valid radar points, an edge case the paper does not stress-test.","The ZJU ground-truth correction implies that published comparisons on that dataset may have been distorted by extrinsic misalignment; if the correction becomes standard, existing baseline numbers could shift, altering leaderboards on that benchmark.","A testable extension is to feed the learned confidence and displacement heads with raw radar from a third sensor type and measure whether recalibration transfers without retraining, which would reveal whether the model learns radar-specific artifacts or generic misprojection patterns."],"forward_implications":["Radar-camera depth estimation can be trained with single-frame sparse LiDAR, cutting supervision density to roughly 1% and removing the need for multi-frame LiDAR accumulation and interpolation.","Refined radar anchors convert any frozen monocular inverse-depth network into metric depth through test-time optimization, so metric accuracy can be added to relative-depth models without retraining them.","The radar recalibration transfers across different radar and LiDAR hardware, as demonstrated by the move from the ZJU platform to the proprietary ZF FRGen21 platform.","Depth maps with accurate scale improve downstream 3D reconstruction quality compared to dense-supervised depth, as shown with the FrozenRecon reconstruction framework.","On 3D radar data lacking elevation information the method is not directly applicable, but the paper notes that elevation inference and multi-view consistency could extend it to such sensors."],"supporting_citations":[{"why":"Radarcam-depth, the main dense-supervised baseline; provides the ZJU-4DRadarCam dataset, the evaluation protocol, and the strongest comparison numbers.","marker":"[9]"},{"why":"Confidence-aware gated radar-image fusion baseline that still needs up to 161 accumulated LiDAR frames during training; RaCalNet's sparse-supervision advantage is measured against it.","marker":"[11]"},{"why":"Deep ordinal regression radar baseline included in the comparison tables.","marker":"[28]"},{"why":"The frozen monocular depth model whose inverse-depth output is aligned; supplies the scale-ambiguous structure that the optimizer rescales.","marker":"[35]"},{"why":"Vision Transformer architecture on which the frozen monocular model is based; standard ingredient of the backbone.","marker":"[36]"},{"why":"Prior metric alignment approach that globally fits monocular predictions to sparse VIO measurements; conceptual precedent for the global affine alignment stage.","marker":"[16]"},{"why":"Multi-sweep radar accumulation with dense LiDAR supervision that RaCalNet explicitly avoids; frames the data-cost problem.","marker":"[29]"}],"fun_headline_variants":["1% LiDAR supervision yields radar depth rivaling dense labels","Radar recalibration makes 1% sparse LiDAR enough for metric depth","Sparse-supervised radar depth cuts RMSE by 35% in real world","Radar depth from 1% sparse LiDAR matches dense-supervised quality","RaCalNet: recalibrated radar points set scale for monocular depth"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The affine-inverse-depth assumption: the frozen monocular network's inverse depth is assumed to be linearly related to true metric inverse depth, both across the whole image and within each cluster; if the network's errors are not affine, the least-squares alignment will bias the predicted metric depth.","fun_headline_variants_meta":{"raw":{"variants":["1% LiDAR supervision yields radar depth rivaling dense labels","Radar recalibration makes 1% sparse LiDAR enough for metric depth","Sparse-supervised radar depth cuts RMSE by 35% in real world","Radar depth from 1% sparse LiDAR matches dense-supervised quality","RaCalNet: recalibrated radar points set scale for monocular depth"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000658,"raw_usage":{"total_tokens":3076,"prompt_tokens":1075,"completion_tokens":2001,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":691,"completion_tokens_details":{"reasoning_tokens":1902}},"tokens_in":691,"tokens_out":2001,"duration_ms":16633,"temperature":1.0,"reasoning_tokens":1902,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:54:00.323272+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render or capture a test scene with wide depth variation concentrated in a single K-means cluster (e.g., a long corridor seen at an angle), run the metric depth optimization with and without per-cluster affine fitting, and measure whether the cluster-wise residual error follows the affine model; if residuals are systematically biased by depth, the affine assumption fails.","supporting_citations":[{"cited_title":"Radarcam-depth: Radar-camera fusion for depth estimation with learned metric scale,","cited_arxiv_id":null,"evidence_quote":"Radarcam-depth, the main dense-supervised baseline; provides the ZJU-4DRadarCam dataset, the evaluation protocol, and the strongest comparison numbers."},{"cited_title":"Depth estimation from camera image and mmwave radar point cloud,","cited_arxiv_id":null,"evidence_quote":"Confidence-aware gated radar-image fusion baseline that still needs up to 161 accumulated LiDAR frames during training; RaCalNet's sparse-supervision advantage is measured against it."},{"cited_title":"Depth estimation from monocular images and sparse radar using deep ordinal regression network,","cited_arxiv_id":null,"evidence_quote":"Deep ordinal regression radar baseline included in the comparison tables."},{"cited_title":"Monocular visual-inertial depth estimation,","cited_arxiv_id":null,"evidence_quote":"Prior metric alignment approach that globally fits monocular predictions to sparse VIO measurements; conceptual precedent for the global affine alignment stage."},{"cited_title":"Radar-camera pixel depth association for depth completion,","cited_arxiv_id":null,"evidence_quote":"Multi-sweep radar accumulation with dense LiDAR supervision that RaCalNet explicitly avoids; frames the data-cost problem."}],"review_version":1}