{"id":"e9fe6cdf-b35e-4af6-a944-437daec3f79e","arxiv_id":"2507.16224","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"LDRFusion reports a LiDAR-dominant two-stage fusion architecture that improves 3D detection on KITTI and nuScenes by refining LiDAR proposals with pseudo point clouds.","lead":"LDRFusion is a two-stage 3D object detection framework that keeps LiDAR in charge of localizing objects, then uses camera-derived pseudo point clouds to refine the boxes, and merges both stages' outputs. It reports small but consistent gains over several strong detectors on the KITTI and nuScenes benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains over the SFD baseline are small (0.19–0.73 mAP on validation for the main ablation) and are presented without error bars, code, or stated HPR hyperparameters, so the central claim depends on the unverified premise that these deltas exceed run-to-run variance.","rationale":"The reader's weakest assumption correctly identifies the central risk: the improvements are small and are reported without error bars, code, or complete hyperparameter specifications. This is the single most load-bearing concern because the paper's main quantitative claim is precisely that LDRFusion outperforms SFD by the amounts in Tables 1, 3, and 6. If those deltas are within run-to-run variance, the proposed architecture contributions are not established. My reading of the manuscript did not reveal a more fundamental internal inconsistency: the method is coherent, the ablation story is plausible, and the KITTI validation and test numbers are broadly consistent with each other. Minor reporting issues such as the Table 4 'LPRFusion' label and the small mismatches between the ablation text and Table 6 values reinforce the need for code and repeated runs but do not independently invalidate the claim. The appropriate disposition remains CONDITIONAL: the idea is credible and worth replicating, but acceptance should be conditioned on a reproducibility check with multiple seeds or released code.","tokens_in":11813,"tokens_out":8687,"duration_ms":89855,"concrete_test":"Obtain or reimplement the authors' code and run the Table 6 configurations (a)–(d) with at least five random seeds under identical settings (SFD training configuration, TWISE depth completion, same KITTI split). Report the mean and standard deviation of 3D car AP (R40) for easy/moderate/hard and mAP. Then check whether the deltas (b)−(a), (c)−(a), and (d)−(a) exceed the inter-seed standard deviation of the respective configurations. Also verify that the reproduced SFD baseline matches the published 95.47/88.56/85.74 within roughly 0.2 mAP; if the baseline shifts by more than that, the reported improvements are not reliable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim is that LDRFusion consistently improves on SFD: on KITTI test, 3D car AP (R40) improves by 0.19/0.71/2.51 points (Table 1), and on the validation set the full method gains 0.53/0.50/0.73 points over the SFD baseline (Table 6, rows (a) vs (d)). The component-wise gains are also small: refinement alone adds 0.26/0.37/0.65 points, and HPR alone adds 0.26 mAP. These deltas are the entire evidentiary basis for the architecture's contributions. The paper reports no standard deviations or repeats, provides no code, and leaves the HPR neighbor-search hyperparameters K and n unspecified (Section 3.3). On the 3769-frame KITTI validation split, single-seed 3D AP differences below 1 mAP are commonly within training noise, so the observed differences cannot be distinguished from run-to-run variance or minor training configuration changes without additional evidence. The load-bearing assumption is therefore that the reported SFD baseline is stable enough that 0.2–0.7 mAP shifts are meaningful. That assumption is asserted, not demonstrated. The absence of code and the unstated HPR parameters also make independent replication impossible as submitted.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LDRFusion, a two-stage LiDAR-dominant fusion framework for 3D object detection. In the first stage, only raw LiDAR point clouds are used to generate proposals; in the second stage, pseudo point clouds obtained from depth completion are incorporated, and the instance-level outputs of both stages are combined with a weighted fusion at inference. A hierarchical pseudo point residual encoding (HPR) module is introduced to capture local structure in pseudo point clouds. Experiments on KITTI and nuScenes report consistent improvements over the SFD baseline, a multi-class validation mAP of 78.71, and comparable inference speed. Ablations in Tables 6 and 7 support the design choices, but the empirical case is weakened by the absence of error bars, unstated HPR hyperparameters, and an ambiguously defined fusion rule.","tokens_in":12139,"tokens_out":7209,"duration_ms":74982,"significance":"If the reported results are reproducible, the paper makes a useful incremental contribution: a simple, efficient LiDAR-dominant refinement strategy that can be plugged into pseudo-point-based detectors, plus an HPR module that improves local representation. The architecture is coherent and the ablations are informative. However, the central empirical claim rests on single-run AP differences that are mostly below 1 mAP on the KITTI validation set, and several reproducibility-critical details are missing. The paper would be a solid systems-style contribution once the statistical reliability of the gains and the missing implementation details are addressed.","major_comments":[{"comment":"The HPR neighborhood size K and pixel search range n are introduced in §3.3 but their numerical values are never given, neither in the method section nor in §4.2. Since HPR is one of the two main contributions and Table 6 credits it with a 0.26 mAP gain, the unstated parameters make the module irreproducible and prevent an assessment of sensitivity. Please specify K and n, and ideally include a sensitivity study over these values.","section":"§3.3 (HPR module)"},{"comment":"The inference-time fusion rule D = αDL + (1−α)DM is not well defined: DL and DM are sets of bounding-box instances, not numeric vectors, so the weighted sum requires a matching or association rule and a score-merging rule (e.g., IoU matching followed by score weighting, then NMS). The paper does not state how the two instance sets are paired, how duplicate detections are removed, or whether α=0.5 is applied per matched pair. This is central to the claimed LiDAR-dominant refinement and must be specified precisely.","section":"§3.2, Eq. (10)"},{"comment":"All reported AP values are single runs without error bars or repeated-seed statistics. The headline ablation gains in Table 6 are +0.53/+0.50/+0.73 mAP over the SFD baseline, and the HPR-only gain is +0.26 mAP; on the 3769-frame KITTI validation split such differences are typically within run-to-run training variance. The paper should provide multiple seeds (mean ± std) or a paired statistical test to support the central claim that the architectural components, not noise or training-configuration variation, cause the improvements.","section":"§4.4, Tables 6 and 7"},{"comment":"The ablation baselines appear inconsistent. Table 6 (a) reports the SFD baseline as 95.47/88.56/85.74, while Table 7 (II), which uses real+pseudo point clouds in both stages with instance-level fusion, reports 95.26/88.26/85.73. If (II) is meant to be the SFD-style configuration, the two numbers should coincide; if not, the Table 7 comparison lacks a stated baseline. Please clarify the exact configuration of Table 7 (II) and how it relates to Table 6 (a), since the LiDAR-dominant conclusion depends on this comparison.","section":"Tables 6 and 7"}],"minor_comments":[{"comment":"Typographical errors include 'soley' in the abstract, 'genrated' in §1, 'fisrt' in §3.2, and 'strucure' in the Figure 3 caption; please proofread the text.","section":"Throughout"},{"comment":"The refinement loop in Eq. (13) concatenates s1 through st+1 but does not state the total number of iterations T or the initial condition for t; please define the loop bounds explicitly.","section":"§3.3, Eq. (13)"},{"comment":"The loss symbols LL_aux and LC_aux are confusing because the subscripts do not match the modalities they supervise (pseudo point clouds vs. raw point clouds). Please rename or define these terms explicitly.","section":"§3.4, Eq. (14)"},{"comment":"The row 'MVP + 2 stage + LPRFusion' presumably should read 'LDRFusion'; the text describes LDRFusion, so the table label appears to be a typo.","section":"Table 4"},{"comment":"The inference-speed comparison does not state measurement conditions such as hardware, batch size, input resolution, or whether depth completion time is included; please provide these details.","section":"Table 5"},{"comment":"Part-Sensitive Warping (PSW) is used in Eq. (2) but is never defined and no precise reference is given for it; please define the operation or cite the source.","section":"§3.2.1, Eq. (2)"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the empirical gap: the reported gains are small and are not supported by repeated runs or variance estimates, and the missing HPR parameters and ambiguous fusion rule prevent replication. If the authors can supply multi-seed results, specify all hyperparameters, and clarify the fusion matching, the paper would be a publishable contribution. There is some fit concern that the incremental improvement over SFD is modest, but the HPR module and the LiDAR-dominant refinement idea are reusable and of interest to the 3D detection community."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on LDRFusion. The paper proposes a two-stage cascade where stage one is LiDAR-only, stage two adds pseudo point clouds from depth completion, and predictions are merged with a fixed weight. The new piece is the HPR module, which encodes residuals within pixel-neighborhood pseudo point sets. That's a real, if modest, architectural contribution, and the ablation story in Tables 6 and 7 is coherent: refinement helps, HPR helps a little, and plugging pseudo points into both stages hurts, which supports the LiDAR-dominant idea.\n\nThe empirical case is where I'd push back. The headline test-set car AP gains over SFD are 0.19/0.71/2.51, and the validation gains for the full method are 0.53/0.50/0.73 on easy/moderate/hard, with mAP gain 0.59. No error bars, no repeats, no code. For KITTI, single-seed differences below 1 mAP are often within train-run noise, so the load-bearing claim that these deltas reflect the architecture rather than variance is asserted, not demonstrated. The paper also never gives K and n for the HPR neighbor search, which makes replication impossible as submitted. Table 4 has a naming slip ('LPRFusion'), minor but indicative of polishing issues.\n\nI don't think this is a fatal problem. The design is sensible, the ablations are in the right direction, and the hard-difficulty gain on test is non-trivial. But the evidence quality is below the bar for acceptance as-is. I'd want the authors to release code, report multiple seeds with standard deviation, and specify all hyperparameters. If the gains survive that, the paper is a solid incremental contribution to pseudo-point fusion. If they don't, it's another sub-1-mAP KITTI paper.\n\nBottom line: worth a serious referee, but my own verdict is conditional. I'd send it to review with a request for code and error bars before acceptance. Not a desk reject, not a strong accept.","headline":"A LiDAR-dominant cascade with a plausible HPR module, but the gains are small, unreplicated, and under-specified, so it deserves conditional review rather than acceptance as-is.","tokens_in":12661,"tokens_out":2155,"would_cite":false,"duration_ms":21967,"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":"LDRFusion claims a LiDAR-dominant two-stage refinement scheme improves 3D detection by adding camera pseudo points only in the second stage, reporting 85.94 mAP on KITTI.","keywords":["LiDAR-camera fusion","3D object detection","pseudo point cloud","depth completion","cascade refinement","residual encoding","KITTI benchmark","autonomous driving"],"falsifier":"Repeat the KITTI car experiment with the same training configuration for both SFD and LDRFusion across ten random seeds; if the 95% confidence interval for the moderate-difficulty AP difference includes zero, the central claim of consistent improvement is not supported.","tokens_in":11622,"feed_emoji":"🚗","tokens_out":6014,"duration_ms":52789,"temperature":0.7,"pith_summary":"LDRFusion is a LiDAR-dominant two-stage refinement framework for 3D object detection. It first generates proposals from LiDAR alone, then refines them with dense pseudo point clouds produced by depth completion, and finally merges the two stages' instance-level predictions. The paper reports that this asymmetric cascade outperforms the SFD baseline on KITTI at every difficulty, with the largest gain on hard cases, and sets a multi-class validation mAP of 78.71. The point is that noisy pseudo points should be introduced only at a refinement stage, where their density helps, rather than fused symmetrically from the start. If correct, the design offers a computationally cheap way to upgrade pseudo-point detectors without retraining from scratch.","feed_headline":"LiDAR-first two-stage fusion lifts 3D car detection to 85.94 mAP","feed_subtitle":"Refining from LiDAR alone, then adding camera pseudo points, beats SFD at every difficulty level.","key_machinery":"The load-bearing object is the hierarchical pseudo point residual encoding (HPR) module. For each pseudo point in a cropped region of interest, HPR finds K neighboring pseudo points by searching nearby image pixels, then iteratively updates each point's feature by concatenating its feature residual relative to the centroid with its own feature and reweighting the result by an MLP applied to the positional residual; aggregation uses an MLP rather than max pooling because RGB features break rotational invariance. This module supplies the second stage's pseudo-point features, while the cascade itself -- LiDAR-only first stage, fused second stage, and weighted instance-level merging $D = \\alpha D^L + (1-\\alpha) D^M$ with $\\alpha = 0.5$ -- is the architectural mechanism the paper credits for the gains.","core_discovery":"The central claim is that detection accuracy improves when the fusion is ordered rather than symmetric: a LiDAR-only stage first produces well-localized proposals, a second stage fuses those proposals with pseudo point clouds to catch hard instances, and the two sets of instance-level outputs are combined with a fixed weight (alpha equal to 0.5). The paper reports KITTI test 3D car AP (R40) of 91.92/85.47/80.43 for easy/moderate/hard (mAP 85.94), compared with SFD's 91.73/84.76/77.92, and a multi-class KITTI validation mAP of 78.71, a 0.92-point improvement over the prior best. It also reports nuScenes mAP/NDS of 67.8/71.0, surpassing the two-stage MVP baseline, and an inference speed of 10.0 FPS, only 0.2 FPS below SFD. The paper argues these results show LiDAR should remain the localization authority while camera-derived pseudo points act as a second-stage enhancer.","pith_inferences":["If the reported KITTI gains hold under repeated-seed evaluation, the same LiDAR-dominant cascade should transfer to other pseudo-point detectors, not just depth-completion ones, because the design only assumes a noisy dense auxiliary cloud.","The class asymmetry (largest gains on cars, second-best on pedestrians and cyclists) suggests the refinement helps most where depth estimation is reliable; a testable extension is to gate the second stage by estimated pseudo-point confidence per class.","Unreported HPR neighborhood parameters (K and the search radius n) mean the method's sensitivity to those choices is unknown; measuring performance across K and n would clarify how much of the gain is architectural versus tuned."],"forward_implications":["On the KITTI test set, the framework reports 3D car AP (R40) of 91.92/85.47/80.43 at easy/moderate/hard, an mAP of 85.94, beating the SFD baseline by 0.19, 0.71, and 2.51 points.","Because the largest improvement is on hard examples, the design specifically helps where sparse LiDAR is weakest.","The nuScenes result (mAP 67.8, NDS 71.0, surpassing the two-stage MVP baseline) indicates the refinement can be attached to other pseudo-point detectors.","Keeping pseudo points out of the first stage avoids injecting depth-completion noise into proposal localization, while instance-level fusion keeps inference near 10 FPS, only 0.2 FPS below SFD.","On joint multi-class validation, the method reaches 78.71 mAP, a 0.92-point gain over the previous best, but the improvement concentrates on cars rather than pedestrians and cyclists."],"supporting_citations":[{"why":"SFD is the multimodal baseline and training-configuration source; the paper's main test-set comparison is measured against it.","marker":"[19]"},{"why":"Voxel R-CNN supplies the voxel pooling and RoI feature extraction used in both stages.","marker":"[10]"},{"why":"The pre-trained depth completion network that generates the pseudo point clouds.","marker":"[37]"},{"why":"CasA provides the cascade attention and feature-level fusion alternative tested in the refine-policy ablation.","marker":"[38]"},{"why":"MVP is the two-stage pseudo-point baseline surpassed on nuScenes.","marker":"[30]"},{"why":"SQD is a recent pseudo-point detector included in the comparisons.","marker":"[22]"},{"why":"PointNet supplies the point-wise MLP encoding idea and the max-pooling aggregation that the HPR module replaces.","marker":"[5]"},{"why":"DGCNN motivates the feature-residual encoding used in HPR.","marker":"[33]"},{"why":"KITTI supplies the dataset and the AP(R40)/AP(R11) evaluation protocol.","marker":"[35]"},{"why":"nuScenes supplies the large-scale multimodal benchmark for the generalization experiment.","marker":"[36]"}],"fun_headline_variants":["LiDAR-first fusion lifts KITTI mAP to 85.94","Two-stage order: LiDAR first, then camera pseudo points","Ordered fusion beats symmetric sensor mixing on KITTI","LiDAR-dominant refinement reaches 85.94 KITTI mAP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reported gains come from the proposed design and not from run-to-run randomness or hidden tuning, since the paper reports no error bars and leaves some HPR hyperparameters (K and the search radius n) unspecified.","fun_headline_variants_meta":{"raw":{"variants":["LiDAR-first fusion lifts KITTI mAP to 85.94","Two-stage order: LiDAR first, then camera pseudo points","Ordered fusion beats symmetric sensor mixing on KITTI","LiDAR-dominant refinement reaches 85.94 KITTI mAP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000323,"raw_usage":{"total_tokens":1822,"prompt_tokens":958,"completion_tokens":864,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":788}},"tokens_in":574,"tokens_out":864,"duration_ms":8305,"temperature":1.0,"reasoning_tokens":788,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:14:56.348119+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the KITTI car experiment with the same training configuration for both SFD and LDRFusion across ten random seeds; if the 95% confidence interval for the moderate-difficulty AP difference includes zero, the central claim of consistent improvement is not supported.","supporting_citations":[{"cited_title":"Sparse fuse dense: Towards high quality 3d detection with depth completion","cited_arxiv_id":null,"evidence_quote":"SFD is the multimodal baseline and training-configuration source; the paper's main test-set comparison is measured against it."},{"cited_title":"V oxel r-cnn: Towards high performance voxel-based 3d object detection","cited_arxiv_id":null,"evidence_quote":"Voxel R-CNN supplies the voxel pooling and RoI feature extraction used in both stages."},{"cited_title":"Depth completion with twin surface extrapolation at occlusion boundaries","cited_arxiv_id":null,"evidence_quote":"The pre-trained depth completion network that generates the pseudo point clouds."},{"cited_title":"Casa: A cascade attention network for 3-d object detection from lidar point clouds.IEEE Transac- tions on Geoscience and Remote Sensing, 60:1–11, 2022","cited_arxiv_id":null,"evidence_quote":"CasA provides the cascade attention and feature-level fusion alternative tested in the refine-policy ablation."},{"cited_title":"Mul- timodal virtual point 3d detection","cited_arxiv_id":null,"evidence_quote":"MVP is the two-stage pseudo-point baseline surpassed on nuScenes."},{"cited_title":"Sparse query dense: Enhancing 3d object detection with pseudo points","cited_arxiv_id":null,"evidence_quote":"SQD is a recent pseudo-point detector included in the comparisons."},{"cited_title":"Pointnet: Deep learning on point sets for 3d classification and segmentation","cited_arxiv_id":null,"evidence_quote":"PointNet supplies the point-wise MLP encoding idea and the max-pooling aggregation that the HPR module replaces."},{"cited_title":"Dgcnn: A convolutional neural network over large-scale labeled graphs","cited_arxiv_id":null,"evidence_quote":"DGCNN motivates the feature-residual encoding used in HPR."},{"cited_title":"Vision meets robotics: The kitti dataset","cited_arxiv_id":null,"evidence_quote":"KITTI supplies the dataset and the AP(R40)/AP(R11) evaluation protocol."},{"cited_title":"nuscenes: A multi- modal dataset for autonomous driving","cited_arxiv_id":null,"evidence_quote":"nuScenes supplies the large-scale multimodal benchmark for the generalization experiment."}],"review_version":1}