{"id":"e391085c-94a0-47fd-81da-a3991d7d2e0c","arxiv_id":"2504.18773","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"CenterDepth estimates object depth from keypoint centers with a local CRF, and VirDepth is a CARLA-based synthetic dataset, but methodological flaws undermine the claimed state-of-the-art results.","lead":"This paper introduces a virtual dataset for autonomous driving and a lightweight network that estimates the distance to cars and pedestrians from the object's center point. But the paper's central module is under-specified and the numbers are internally inconsistent, so the headline results are not yet supported.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Center FC-CRFs module is not specified tightly enough to be trainable or to explain the reported gains; the central claim rests on an unverifiable component.","rationale":"The reader's weakest assumption identifies the same load-bearing issue: Center FC-CRFs is assumed to be a well-defined, trainable component, but the equations and surrounding text do not specify it. I agree with that assessment. The strongest claim is unverifiable from the manuscript because the mechanism to which all major gains are attributed is not implementable as written. The internal metric inconsistencies across Tables I–III provide additional evidence that the numerical results are not reliable enough to support an order-of-magnitude improvement. Since the paper offers no formal verification, no working code, and no parameter-free derivation to independently vouch for the module, the central claim cannot be accepted from the evidence presented. My critique does not move the verdict because the reader already recommended REJECT; I would keep that verdict.","tokens_in":14383,"tokens_out":3564,"duration_ms":40191,"concrete_test":"Obtain the released code and checkpoints, then run the exact CenterDepth versus CenterDepth* comparison on virDepth with identical training data, backbone initialization, schedule, and evaluation protocol, toggling only the Center FC-CRFs module as defined by the manuscript. If the code is unavailable, independently re-implement the module from Eqs. 2–5 alone: if H, ψu, xε, the training loss, and the inference minimization cannot be instantiated unambiguously, the manuscript as written does not support the central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that CenterDepth's Center FC-CRFs module produces large accuracy gains—depends entirely on the module being well-defined and on the ablations isolating it. Neither condition is met. In Section III.B, Eqs. 2–5 use undeclared symbols: Eq. 2 defines (xc,yc) as arg max of H(x,y), while the text says B(x,y) is the heatmap; H is never defined. Eq. 5 introduces xε without definition, and the unary potential ψu is not given. No loss function, training objective, or inference procedure for the depth D is specified; the text only says the final depth is obtained by minimizing E(D), but the network outputs, the differentiable surrogate, and the optimization algorithm are absent. The comparisons in Tables I, II, and IV attribute large improvements—e.g., virDepth HGL104 δ1 from 0.804 to 0.989 and R4 MAE from 13.537 to 3.351—to this module, yet CenterDepth* is never defined beyond appearing to be a baseline. The reported numbers are also internally inconsistent: Table I gives HGL104 MRE=0.041 while Table III reports 1.13%, and ResNet-101 MRE differs between Table II (0.053) and Table III (2.31%) for the same configuration. Without a complete module specification and a controlled ablation, the headline result cannot be reproduced or attributed to the proposed mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two contributions: a CARLA-based automated pipeline for generating a synthetic autonomous-driving dataset (VirDepth) with synchronized RGB, depth, semantic, and 3D-box annotations, and a monocular object-centric depth estimation framework called CenterDepth. CenterDepth combines a CenterNet-style keypoint detector with a novel 'Center FC-CRFs' module that is claimed to propagate information from the detected object center within the detection window, yielding high accuracy and long-range stability. The authors report strong results on VirDepth, Virtual KITTI 2, KITTI-Depth, and KITTI-3D, including δ1 of 0.989 and MRE of 0.041 for monocular depth on VirDepth, low MAE in the 150-200 m range, and real-time inference on embedded hardware.","tokens_in":14738,"tokens_out":6976,"duration_ms":71005,"significance":"If the reported results are credible, the paper would offer a low-cost way to generate task-specific virtual driving datasets and a lightweight depth estimator that degrades gracefully at long range. The dataset-generation pipeline is concrete and useful in principle, and the paper explicitly evaluates the difficult 100-200 m regime, which is a genuine practical concern. However, the significance cannot currently be assessed: the central Center FC-CRFs module is not specified at a reproducible level, the ablation baseline is undefined, and the numerical results are internally inconsistent across tables. Because the main technical novelty and the headline accuracy gains both rest on this module, the paper in its present form does not support its central claims.","major_comments":[{"comment":"The Center FC-CRFs module, which is the paper's main technical novelty, is not defined precisely enough to be implemented or evaluated. Eq. (2) defines the center as the arg max of H(x,y), while the text says B(x,y) is the heatmap and H is never introduced. Eq. (5) uses xε,yε without definition. The unary potential ψu is never specified, and the text only says that the final depth is obtained by minimizing E(D); no loss function, network output parametrization, differentiable surrogate, or optimization algorithm is given. Since Tables I, II, and IV attribute the large gap between CenterDepth and CenterDepth* to this module, the central claim is not reproducible from the manuscript.","section":"Section III.B, Eqs. (2)-(5)"},{"comment":"The baseline CenterDepth* is used throughout the experimental section as the comparison point for ablation claims, but it is never defined. The text identifies it only as the 'baseline model' in Section V.C; it is not stated whether it is CenterDepth without the Center FC-CRFs module, without the center anchor, or without some other component, nor how it was trained. Consequently, the claimed improvements attributable to the proposed module, such as δ1 improving from 0.804 to 0.989 on VirDepth with HGL104 and R4 MAE improving from 13.537 to 3.351, cannot be verified or assigned a cause.","section":"Section V.C, Tables I, II, IV"},{"comment":"The manuscript contains multiple contradictory quantitative statements that undermine confidence in the reported results. The dataset is described as containing 'more than 20,000 images' generated in four hours, 'more than 50,000 images' across five scenes, 'roughly 15,000 images' in five scenes in Section V.A, and 'eight urban scenes' in the Introduction. For Hourglass-104, MRE is reported as 0.041 in Table I but as 1.13% in Table III; for ResNet-101, MRE is 0.053 in Table II but 2.31% in Table III. Runtime claims in Section V.D (e.g., 0.025 s for ResNet-18 and 0.038 s for Hourglass-104) do not match Table III values of 3.4 ms and 15.6 ms, respectively. These inconsistencies are too large to be explained by unit conventions.","section":"Sections IV, V.A, V.C, V.D and Tables I, III"},{"comment":"The evaluation protocol for full-image depth baselines is underspecified. The paper does not state whether DepthAnythingV2 and MonoDepth2 were fine-tuned on VirDepth, which checkpoint or input resolution was used, or how their output depth scales were aligned to the metric ground truth before applying the Center or Seg extraction. Because CenterDepth is trained on the same distribution it is tested on, the comparison may mix architectural advantage with training-data exposure, and the reported margins over these baselines cannot be interpreted without this information.","section":"Section V.A and V.C"}],"minor_comments":[{"comment":"The index term 'Depth Esitimation' is misspelled; it should be 'Depth Estimation'.","section":"Index Terms"},{"comment":"The notation is inconsistent: Eq. (2) uses H(x,y) while the text identifies B(x,y) as the heatmap; please use one symbol consistently and define both the heatmap and the feature map F.","section":"Section III.B"},{"comment":"The variables xε,yε in Eq. (5) are undefined; they appear to be intended to denote the center point (xc,yc) used in Eq. (3), but this should be stated explicitly.","section":"Section III.B, Eq. (5)"},{"comment":"The heading 'Limitations and Systematic Analysis' is misleading: the subsection contains a path-planning demonstration but no discussion of limitations or a systematic error analysis. Please retitle or expand accordingly.","section":"Section V.E"},{"comment":"Several figure callouts read 'as shown in the figure' without a figure number, and Table V lists Params(M)/FLOPs(B) for CenterDepth that do not match any backbone in Table III; both need clarification.","section":"Section V.E and Table V"}],"recommendation":"reject","confidential_remarks":"For the editor: the paper's central method cannot be inspected because Sections III.B and the experimental protocol omit essential implementation and evaluation details, and the quantitative claims are inconsistent across tables. These are not merely presentation issues: the main contribution, the Center FC-CRFs module, is underspecified, and the ablation baseline is undefined. Even with a full rewrite of the method section and corrected tables, the experiments would need to be redone or substantially revalidated to establish the claimed gains. I therefore recommend rejection rather than major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on arXiv:2504.18773. The most useful thing in this paper is the dataset engineering. The CARLA-based collector that produces aligned RGB/depth/segmentation frames and the VirDepth set with 180k targets spread across 0-200m is a real convenience for people who need scenario-specific synthetic data. I also give them credit for reporting distance-binned MAE and running on a Jetson; that is the right kind of evaluation for an efficiency-focused perception paper.\n\nThe rest of the paper is not in a publishable state. The central module, Center FC-CRFs, is described with equations that don't parse. Eq. (2) takes an argmax over H(x,y) while the text says B(x,y) is the heatmap. Eq. (5) uses xϵ without definition, and ψu is never given. No loss, no inference procedure, no optimization algorithm for E(D). That matters because the reported gains come from comparing CenterDepth against CenterDepth*, and CenterDepth* is never defined. The ablation cannot isolate the module's contribution. The numbers also disagree with each other: VirDepth is 20k, 50k, and roughly 15k images at different points; scenes are 5 and then 8; HGL104 MRE appears as 0.041 and 1.13%; ResNet-101 as 0.053 and 2.31%. Comparison methods appear to be zero-shot rather than retrained on VirDepth, so the headline margins are not apples-to-apples.\n\nThe reader's stress-test note is fair and lands. The circularity worry I would downgrade: this is ordinary supervised regression with a smoothness prior, not self-fulfilling reasoning, and the equations are not the kind of thing that could be circular. But the external validity concern is real because almost everything is evaluated on the authors' own synthetic set.\n\nCan it be fixed? Yes, most of these are exposition and experimental-discipline problems. They need to write down the energy minimization and the training objective, define CenterDepth*, retrain or clearly label the baselines, and make the dataset statistics and metric tables consistent. If they do that, the method may be a useful engineering contribution. As it stands, the central claim cannot be verified from the manuscript.\n\nI would not cite this, and I would not bring it to a reading group. My recommendation: desk-reject as-is, with encouragement to resubmit after the rewrite. It is not worth referee time yet.","headline":"A useful synthetic-data engineering idea is buried under an underspecified core module and inconsistent numbers; the central claim is not verifiable as written.","tokens_in":15225,"tokens_out":4042,"would_cite":false,"duration_ms":45117,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Monocular depth for driving can be estimated from an object's center point using a local conditional random field, the paper claims.","keywords":["monocular depth estimation","center point detection","conditional random fields","autonomous driving","virtual dataset","CARLA","keypoint-based depth","long-range perception"],"falsifier":"Implement Center FC-CRFs exactly from the energy function in Eqs. (2)-(5) and train the same backbone on VirDepth with the module removed versus included; the paper gives no loss term or inference update rule for the energy minimization, so a reader cannot reproduce it. If a direct implementation does not reproduce the large gap between CenterDepth and CenterDepth* reported in Tables I and IV, the central attribution of gains to the module is wrong.","tokens_in":14214,"feed_emoji":"🚗","tokens_out":6878,"duration_ms":62475,"temperature":0.7,"pith_summary":"The paper argues that monocular depth estimation for autonomous driving should be object-centric rather than dense: locate each vehicle or pedestrian by its center point, then regress that center's depth directly. To make this work at long range, it introduces Center FC-CRFs, a fully-connected conditional random field restricted to the detection box, with the center acting as the anchor for feature and depth propagation. The paper also contributes a CARLA-based data generator and the VirDepth synthetic dataset, built in hours with KITTI-matching camera parameters. On VirDepth, Virtual KITTI 2, and KITTI subsets, it reports that CenterDepth reaches $\\delta_1$ 0.989 and MRE 0.041 with the Hourglass-104 backbone and MAE 3.351 m in the 150-200 m band, at real-time inference speeds. If these results hold, the implication is that center-point depth can stay accurate where full-image depth estimators and monocular 3D detectors degrade on small distant targets.","feed_headline":"Center-point depth hits 98.9% accuracy at 200 m range","feed_subtitle":"Object-center regression with a local CRF keeps mean error at 3.35 m for 150-200 m targets, in real time.","key_machinery":"The load-bearing mechanism is the Center FC-CRFs module: a fully-connected conditional random field restricted to the object's bounding box, with the detected center point as the anchor node. Pairwise potentials tie each pixel's depth to the center depth with weights based on feature similarity (Eq. 4), and the depth map is obtained by minimizing the energy in Eq. 5. This module is what converts the center-point heatmap into a depth value per target, and the paper attributes the large gains of CenterDepth over its no-CRF baseline to it. The second mechanism is the virtual data generator: CARLA meshes are projected through KITTI-matched intrinsics and extrinsics to produce 2D boxes, 3D boxes, and dense depth labels automatically, creating VirDepth with more than 180,000 targets between 0 and 200 m.","core_discovery":"CenterDepth's central claim is that depth estimation in driving scenes does not need a per-pixel depth map; it needs the depth of each object. The system predicts a heatmap of object centers, takes the center as a feature anchor, and uses the Center FC-CRFs module to propagate information from the center to all pixels inside the predicted bounding box, so the object's depth is regressed with global semantic context but without global computation. The paper reports that this design yields a $\\delta_1$ of 0.989 and MRE of 0.041 on VirDepth, that it outperforms DepthAnythingV2, MonoDepth2, SMOKE, FCOS3D, DD3D, and BEVFormer on VirDepth, Virtual KITTI 2, and KITTI test sets, and that in the 150-200 m interval its MAE of 3.351 m is far below the 36.269 m of DepthAnythingV2 and 6.354 m of BEVFormer. It further claims the long-range advantage comes specifically from the Center FC-CRFs module, since the same network without it (CenterDepth*) performs markedly worse in the distance-subdivided evaluation.","pith_inferences":["A direct test of the mechanism would replace Center FC-CRFs with a simple attention-weighted average of features around the center; if the gains persist, the energy formulation itself is not the cause.","The reported large-scale balancing is demonstrated inside the synthetic distribution of VirDepth; on real roads, occlusion, weather, and sensor noise at 150-200 m are likely to erode the advantage, so the transfer claim should be read as conditional on the synthetic training distribution.","The same center-anchored CRF idea could be applied to depth completion or to anchor-free 3D detectors, where the center heatmap is already computed and the extra cost of a local CRF would be small.","Because the paper filters out targets beyond 200 m and highly occluded objects, the method's practical ceiling in denser urban traffic is still unknown; an occlusion-stratified evaluation would be a natural next experiment."],"forward_implications":["If the reported accuracy is real, dense pixel-wise depth maps are not required for obstacle localization: detecting the center and regressing center depth can feed BEV projection and path planning at lower compute cost.","Long-range perception beyond 150 m becomes feasible with a single camera, since the reported 150-200 m MAE stays around 3.35 m instead of exploding to tens of meters as it does for full-image estimators.","The speed-accuracy trade-off is controllable by backbone choice: ResNet-18 runs at 3.4 ms and Hourglass-104 at 15.6 ms on an RTX 4090, with ResNet-101 near Hourglass accuracy at 9.5 ms.","Synthetic datasets generated in hours with KITTI-matched calibration can serve as training and evaluation grounds for depth and 3D detection, reducing the cost of task-specific data collection.","Combining the center-point anchor with semantic segmentation and 3D detection objectives in one network can improve both small-object detection (mAP 45.4 vs 30.0 for YOLOv8s in the 150-200 m band) and depth quality."],"supporting_citations":[{"why":"Supplies the center-point detection backbone (Hourglass-104 initialization) that predicts object centers used as depth anchors.","marker":"[27]"},{"why":"Provides the neural window FC-CRFs formulation that Center FC-CRFs adapts to object bounding-box windows.","marker":"[47]"},{"why":"The CARLA simulator used to automate generation of the new VirDepth dataset with calibrated cameras.","marker":"[18]"},{"why":"Sets the camera intrinsics and extrinsics that VirDepth mimics and provides the KITTI benchmark test sets.","marker":"[9]"},{"why":"Public virtual dataset used as a cross-dataset test and as a comparison for the dataset-construction method.","marker":"[48]"},{"why":"Strongest full-image depth estimation baseline; the VirDepth $\\delta_1$ comparison (0.832 vs 0.989) anchors the central accuracy claim.","marker":"[41]"},{"why":"Monocular 3D detection baseline and source of the ResNet-101-DCN initialization used for CenterDepth.","marker":"[39]"},{"why":"Keypoint-based 3D detection baseline compared in the center-depth evaluation tables.","marker":"[40]"}],"fun_headline_variants":["Depth by object center beats per-pixel at 200 m","CenterDepth: object-center regression cuts 200-m error to 3.35 m","Object-center depth with CRF: best long-range, real-time","CenterDepth: 0.989 δ1, beats 6 baselines at all ranges","Depth from object points: 3.35-m error at 150–200 m"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Center FC-CRFs module is a well-defined, trainable component whose described energy minimization is what produces the large accuracy gains, even though the paper leaves its loss and inference procedure unspecified.","fun_headline_variants_meta":{"raw":{"variants":["Depth by object center beats per-pixel at 200 m","CenterDepth: object-center regression cuts 200-m error to 3.35 m","Object-center depth with CRF: best long-range, real-time","CenterDepth: 0.989 δ1, beats 6 baselines at all ranges","Depth from object points: 3.35-m error at 150–200 m"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000394,"raw_usage":{"total_tokens":2069,"prompt_tokens":944,"completion_tokens":1125,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":1022}},"tokens_in":560,"tokens_out":1125,"duration_ms":8156,"temperature":1.0,"reasoning_tokens":1022,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:09:36.063063+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Implement Center FC-CRFs exactly from the energy function in Eqs. (2)-(5) and train the same backbone on VirDepth with the module removed versus included; the paper gives no loss term or inference update rule for the energy minimization, so a reader cannot reproduce it. If a direct implementation does not reproduce the large gap between CenterDepth and CenterDepth* reported in Tables I and IV, the central attribution of gains to the module is wrong.","supporting_citations":[{"cited_title":"Neural window fully- connected crfs for monocular depth estimation,","cited_arxiv_id":null,"evidence_quote":"Provides the neural window FC-CRFs formulation that Center FC-CRFs adapts to object bounding-box windows."},{"cited_title":"Smoke: Single-stage monocular 3d object detection via keypoint estimation,","cited_arxiv_id":null,"evidence_quote":"Keypoint-based 3D detection baseline compared in the center-depth evaluation tables."}],"review_version":1}