{"id":"ca0a59a8-b3a5-400a-8fe3-516cbbfa1978","arxiv_id":"2507.13899","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Adding DepthAnything V2 depth estimates to LiDAR point features raises KITTI 3D detection AP for pedestrians and cyclists by 3.3 and 1.1 points, while dropping cars by 1.0 point.","lead":"LiDAR point clouds give each reflected point four values, and one of them, reflectance, does a poor job telling cars from pedestrians. This paper adds a fifth value, a depth estimate from the pretrained vision model DepthAnything, and shows on KITTI that the extra channel helps detect pedestrians and cyclists, while slightly hurting cars.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Depth-prior attribution is unproven: dDA is unspecified/unnormalized, and the ablations do not isolate it from BGRF or from a trivial range channel.","rationale":"The reader's weakest assumption correctly targets the uncalibrated, underspecified dDA channel. I sharpened this into a concrete causal concern: the only information dDA adds over (x,y,z) is the DepthAnything error field, which is class-correlated on KITTI. The experiment's own sign pattern supports this reading, and the missing BGRF-only row weakens the attribution of the gain to the depth prior. These points are load-bearing because the paper's central claim is that fusing foundation-model depth priors causes the accuracy improvement, not merely that the enriched architecture performs well. The paper's results are plausible and addressable, so the reader's CONDITIONAL verdict remains appropriate; it should not be hard rejected on the basis of these omissions, but the authors should supply the requested control and the missing ablation before the attribution claim is accepted.","tokens_in":11242,"tokens_out":11797,"duration_ms":152247,"concrete_test":"Re-run the Table 3 Pedestrian/Cyclist comparison with the fifth channel replaced by the radial LiDAR range (sqrt(x^2+y^2+z^2)), normalized to the same mean and standard deviation as dDA, while keeping all other settings and the training schedule identical; also include the missing BGRF-only ablation row from Table 4. If the radial-range control reproduces most of the +3.33/+1.12 gains, or if the BGRF-only row already matches the with-depth mAP, then the reported improvement is not caused by DepthAnything's depth prior and the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1 concatenates a DepthAnything V2 depth sample dDA with (x,y,z,r) but never states whether the relative or metric variant is used, nor how dDA is normalized before concatenation. This is not merely a reproducibility gap: because the LiDAR point projects to the same surface, dDA is largely a noisy copy of the LiDAR's own range, so the only new information in the fifth channel is the monocular depth estimator's error pattern. DepthAnything V2 errors on KITTI are correlated with object scale (pedestrians and cyclists are commonly merged into background), so the network may learn a class-correlated residual (dDA - z) rather than a geometric prior. The Table 3 pattern of +3.33 Pedestrian, +1.12 Cyclist, and -1.05 Car is exactly what such an artifact predicts, and the paper's 'misalignment' explanation is not tested. Additionally, Table 4's text references a BGRF-only ablation row, but the table omits that row, so the depth-prior effect cannot be cleanly separated from the fusion module.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper argues that augmenting LiDAR point clouds with depth values predicted by DepthAnything V2 improves 3D object detection. The proposed pipeline concatenates the sampled monocular depth dDA with (x,y,z,r), feeds the resulting five-dimensional points through a PointGFE module and a dual-path RoI extraction framework (voxel-based RoI Grid Pooling and point-based RoI Aware Pooling), and merges the RoI features with a bidirectional gated fusion module (BGRF). On KITTI, the authors report gains over their no-depth-prior baseline for Pedestrian (+3.33 AP) and Cyclist (+1.12 AP) and a drop for Car (-1.05 AP), with a Car-focused ablation showing that DPL+BGRF improves over the baseline. The paper acknowledges the Car drop in Section 4.3 and the conclusion, but the abstract and contribution list claim consistent improvements across categories.","tokens_in":11515,"tokens_out":8997,"duration_ms":102562,"significance":"If the depth-prior benefit is real and attributable to the DepthAnything feature itself, this is a practical and low-cost way to inject foundation-model geometry into LiDAR detectors, with clear value for small or partially occluded objects. The paper reports a complete multi-class KITTI evaluation, a reasonable runtime analysis, and an ablation that at least shows the combined system helps on Car. The contribution is, however, conditional on resolving three concerns: the depth channel is under-specified, no ablation isolates the monocular depth value from a trivial LiDAR-derived range channel, and an ablation row needed to separate DPL from BGRF is missing. These issues are fixable within the scope of the paper, but until they are addressed the central attribution claim is not established.","major_comments":[{"comment":"The manuscript does not state which DepthAnything V2 model is used (metric or relative depth), what preprocessing/normalization is applied to dDA before concatenation with (x,y,z,r), or how the value is sampled at projected LiDAR coordinates (nearest neighbor, bilinear, occlusion handling). Because the entire reported gain is attributed to this fifth channel, the experiment is not reproducible and the effect cannot be interpreted without this information. Please specify the variant/checkpoint, the normalization scheme, and the sampling procedure, and include a sensitivity check on normalization.","section":"Section 3.1, Table 3"},{"comment":"The paper does not test whether a trivial, LiDAR-native depth channel would produce the same gains. Since dDA is sampled at the projection of each LiDAR point, it is largely a re-estimate of the same surface depth; the incremental information is mostly the monocular estimator's error pattern, which on KITTI is correlated with object scale and category. The observed pattern (+3.33 Pedestrian, +1.12 Cyclist, -1.05 Car) is consistent with the network learning a class-correlated residual rather than a geometric prior. Please add ablations that replace dDA with (i) a LiDAR-derived range or camera-frame depth coordinate and (ii) the same coordinate with calibration-like noise or arbitrary scaling, to determine whether the gains come from the extra dimension or from the specific DepthAnything output.","section":"Section 3.1, Table 3"},{"comment":"The ablation text states that \"When only BGRF is introduced, a more notable gain is observed (85.05 vs. 86.13),\" but Table 4 contains no BGRF-only row; the 86.13 result corresponds to DPL+BGRF. This makes it impossible to separate the contribution of BGRF from that of DPL and does not support the claimed evaluation of individual and combined effects. Add the missing BGRF-only row or rewrite the paragraph so it describes the actual table entries.","section":"Section 4.4, Table 4"},{"comment":"The abstract and the contribution bullet claim \"consistent performance improvements across Car, Pedestrian, and Cyclist categories\" (and the abstract says the method \"consistently improves detection accuracy\"), but Table 3 reports a 1.05 AP drop for Car. Although Section 4.3 and the conclusion acknowledge this drop, the summary claims are contradictory and should be revised to state the mixed outcome (gains on Pedestrian/Cyclist, slight loss on Car).","section":"Abstract, Section 1 contributions, Section 4.3"}],"minor_comments":[{"comment":"Report multiple runs or error bars for the w/ and w/o depth-prior conditions; the +1.12 Cyclist gain is small enough that single-run AP values could be within run-to-run variation.","section":"Table 3"},{"comment":"The caption contains garbled text (\"Laplacemub Car 0 0.13198 ...\") that appears to be fit parameters; please clarify the notation and what distribution is being fitted.","section":"Figure 1"},{"comment":"The phrase \"independent of the original LiDAR measurements\" is misleading because the depth map is derived from an image of the same scene; rephrase as \"derived from a different sensor modality\" or similar.","section":"Section 3.1"},{"comment":"The implementation details omit the training schedule (epochs, batch size, learning rate, and any changes relative to the Voxel R-CNN baseline); adding these is necessary for reproducibility.","section":"Section 4.2"},{"comment":"State the hardware and settings for the 9.8 FPS runtime and clarify that the component timings include the cost of all three cascade stages, as the text appears to already intend.","section":"Section 4.5"},{"comment":"The module called \"Depth Prior Learning (DPL)\" in the ablation is called \"depth prior augmentation\" in the method; use consistent terminology.","section":"Sections 3.1 and 4.4"},{"comment":"The column heading \"MAP\" should be \"mAP\" for consistency with the text.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the manuscript is within ITSC scope and the idea is timely, but the missing controls and unspecified depth-channel details make the central claim unverified. I recommend major revision; acceptance should require the additional ablations and clarifications described in the report."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is an incremental but plausible paper. The one genuinely new piece is using DepthAnything V2 as a drop-in depth prior for LiDAR points—prior work used depth completion trained on the target dataset. The reported gains are real to the eye: +3.33 AP on pedestrians, +1.12 on cyclists, with -1.05 on cars. Table 3, if the baseline is the full model without dDA, cleanly isolates the depth prior's contribution, so the stress-test's claim that the ablation doesn't separate it from BGRF is off for the main result. The missing BGRF-only row is in Table 4, the Car-only ablation, and the text referring to it is an error.\n\nWhat the paper does well: it is honest enough to report the Car drop and offer a plausible explanation; the architecture parts (PointGFE, BGRF) are standard but the gated fusion is described adequately; the runtime breakdown is useful.\n\nSoft spots: the paper never specifies which DepthAnything V2 checkpoint (metric vs relative depth) it uses, nor how dDA is scaled before concatenating with (x,y,z,r). That is the load-bearing detail for reproducibility, and it feeds the concern that the network is learning the error pattern (dDA - z) as a class cue. I do not think the paper's results prove that artifact, but the authors need to rule it out, e.g., by adding a trivial range-channel ablation. Also missing: error bars, code, and a single dataset only. The abstract says 'consistent improvements' which the Car drop contradicts; the body admits it, so fix the wording.\n\nWho it is for: people working on LiDAR detection fusion who want a quick empirical data point. Fine for a workshop or a conference short paper after revision.\n\nMy recommendation: send to peer review. The claim is testable, the missing details are fixable, and the core result—if it survives a check of the depth specification—would be a useful data point.","headline":"A plausible but under-specified depth-prior augmentation for LiDAR detection; Table 3 supports the core claim, but missing depth details and an inconsistent Table 4 row need fixing before I'd trust the gains.","tokens_in":12024,"tokens_out":3670,"would_cite":false,"duration_ms":41672,"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":"The paper claims that appending a DepthAnything V2 depth value to each LiDAR point improves 3D detection for pedestrians and cyclists on KITTI, at a small cost for cars.","keywords":["3D object detection","LiDAR point cloud","depth prior","foundation model","monocular depth estimation","depth feature fusion","KITTI benchmark"],"falsifier":"Run the identical depth-prior augmentation on a second LiDAR-camera dataset with different sensor and calibration characteristics, and compare pedestrian and cyclist AP with and without the prior; if the gains vanish or reverse, the central claim that foundation depth priors reliably improve LiDAR detection fails. A complementary check is to measure, per class, the correlation between $d_{\\mathrm{DA}}$ and LiDAR ray distance: if the correlation differs sharply by class rather than by range, the network is exploiting a class artifact, not geometry.","tokens_in":11100,"feed_emoji":"🚗","tokens_out":7413,"duration_ms":75161,"temperature":0.7,"pith_summary":"The paper claims that raw LiDAR point features—$(x,y,z)$ coordinates plus reflectance $r$—are not expressive enough for reliable 3D object detection, and that this can be fixed by appending a depth value $d_{\\mathrm{DA}}$ predicted by the monocular foundation model DepthAnything V2 at the image projection of each LiDAR point. The resulting five-dimensional feature $(x,y,z,r,d_{\\mathrm{DA}})$ is fed through a point-wise feature extractor, a dual-path RoI extraction with voxel and point branches, and a bidirectional gated fusion module. On the KITTI validation set this yields +3.33 AP for Pedestrian and +1.12 AP for Cyclist, while Car drops by 1.05 AP, a trade-off the authors attribute to misalignment between image-predicted depth and LiDAR geometry on large objects. If the claim is right, pre-trained monocular depth models can serve as a drop-in geometric prior for LiDAR perception, avoiding dataset-specific depth completion.","feed_headline":"Depth priors lift LiDAR pedestrian detection by 3.33 AP","feed_subtitle":"Adding DepthAnything V2 depth to LiDAR points sharpens small-object detection; cars need selective fusion instead.","key_machinery":"The central object is the enriched LiDAR point $(x,y,z,r,d_{\\mathrm{DA}})$, where $d_{\\mathrm{DA}}$ is the depth value sampled from the DepthAnything V2 depth map at the pixel where a LiDAR point projects via known camera-LiDAR calibration. Around it the paper builds three mechanisms: PointGFE, a point-wise extractor that combines point features with ball-query relative positions; a dual-path RoI branch pairing RoI Grid Pooling (global voxel context) with RoI Aware Pooling (fine-grained point structure); and a bidirectional gated RoI fusion module (BGRF) that uses learned attention gates across three cascaded stages. The mechanism's job is to let the depth-enriched point branch compensate for information lost in voxelization while the gating keeps global and local features balanced.","core_discovery":"On the paper's own terms, the central discovery is that a foundation-model depth prior adds discriminative geometry to LiDAR points precisely where reflectance fails: the reflectance distributions of Pedestrian and Cyclist overlap heavily, while DepthAnything V2 separates objects by distance. Concatenating the sampled depth value with the raw attributes gives the detector a cue that is independent of surface material and incidence angle, and the proposed architecture is built to keep that cue alive through both voxel and point branches. The full method reaches 93.12/86.13/83.65 AP for Car on the KITTI validation set and 64.96 mean AP on the official test set, with the multi-class validation comparison showing the pedestrian and cyclist gains. Table 3 is the evidence for the core claim: with versus without depth priors, mAP rises from 56.34 to 59.67 for Pedestrian and from 72.28 to 73.40 for Cyclist, and falls from 88.59 to 87.54 for Car.","pith_inferences":["My inference: the pedestrian/cyclist gains may come less from geometric accuracy of $d_{\\mathrm{DA}}$ than from class-correlated cues—DepthAnything's monocular depth encodes object scale, silhouette, and distance patterns that correlate with class—so the method's transfer to new sensors or city distributions is not guaranteed by the KITTI numbers alone.","My inference: a selective gating rule based on LiDAR-depth consistency (e.g., disable the prior where $d_{\\mathrm{DA}}$ disagrees with the ray distance) could plausibly recover the Car loss while retaining the pedestrian gain; this is testable with the authors' own ablation setup.","My inference: because the enrichment is just an extra input channel, the same recipe could be dropped into other point- or voxel-based detectors, making the paper's contribution a reusable depth-channel prior rather than a detector-specific fix."],"forward_implications":["Depth priors from foundation models can be inserted into LiDAR detectors without dataset-specific depth completion or fine-tuning of the depth model.","Categories with weakly discriminative reflectance, notably pedestrians and cyclists, are where the prior helps most; cars, where LiDAR geometry is already strong, do not benefit.","The added modules are cheap in runtime—PointGFE costs 0.021 s, RoI Aware Pooling 0.006 s, and BGRF 0.013 s per frame—so the prior pipeline stays close to real time at 9.8 FPS.","RoI Grid Pooling, inherited from Voxel R-CNN, remains the computational bottleneck at 0.044 s, pointing to a natural target for future speedups.","The Car AP drop indicates uniform fusion of predicted depth is not always beneficial; adaptive or scale-aware prior integration is the stated direction for future work."],"supporting_citations":[{"why":"Supplies the DepthAnything monocular depth foundation model whose predicted depth motivates the prior.","marker":"[5]"},{"why":"Supplies DepthAnything V2, the model actually used to produce the depth values concatenated into the point features.","marker":"[6]"},{"why":"Establishes the Voxel R-CNN baseline, the RoI Grid Pooling branch, and the proposals the framework builds on.","marker":"[16]"},{"why":"Provides RoI Aware Pooling used in the point-based branch to aggregate fine-grained local features.","marker":"[17]"},{"why":"Provides the KITTI dataset and evaluation protocol on which every reported AP number is measured.","marker":"[26]"},{"why":"Serves as the cross-modal point-prior baseline and motivation that other priors can enrich LiDAR points.","marker":"[20]"},{"why":"Supplies the open-source implementation base used for training and evaluation.","marker":"[27]"}],"fun_headline_variants":["Depth priors boost LiDAR pedestrian and cyclist detection","Foundation model depth priors sharpen LiDAR small-object detection","Depth priors lift LiDAR detection for pedestrians and cyclists","Visual foundation priors enhance LiDAR 3D detection for small objects"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the depth value sampled from DepthAnything V2 at each projected LiDAR point is a useful, properly scaled geometric cue; the paper neither identifies which DepthAnything V2 variant it uses nor states how $d_{\\mathrm{DA}}$ is normalized, so if the input is uncalibrated relative depth, the gains may be a KITTI-specific correlation with object class rather than transferable geometry.","fun_headline_variants_meta":{"raw":{"variants":["Depth priors boost LiDAR pedestrian and cyclist detection","Foundation model depth priors sharpen LiDAR small-object detection","Depth priors lift LiDAR detection for pedestrians and cyclists","Visual foundation priors enhance LiDAR 3D detection for small objects"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000568,"raw_usage":{"total_tokens":2701,"prompt_tokens":968,"completion_tokens":1733,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":1663}},"tokens_in":584,"tokens_out":1733,"duration_ms":15041,"temperature":1.0,"reasoning_tokens":1663,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:13:45.936568+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical depth-prior augmentation on a second LiDAR-camera dataset with different sensor and calibration characteristics, and compare pedestrian and cyclist AP with and without the prior; if the gains vanish or reverse, the central claim that foundation depth priors reliably improve LiDAR detection fails. A complementary check is to measure, per class, the correlation between $d_{\\mathrm{DA}}$ and LiDAR ray distance: if the correlation differs sharply by class rather than by range, the network is exploiting a class artifact, not geometry.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies DepthAnything V2, the model actually used to produce the depth values concatenated into the point features."},{"cited_title":"V oxel r-cnn: To- wards high performance voxel-based 3d object detec- tion","cited_arxiv_id":null,"evidence_quote":"Establishes the Voxel R-CNN baseline, the RoI Grid Pooling branch, and the proposals the framework builds on."},{"cited_title":"From points to parts: 3d object detection from point cloud with part-aware and part- aggregation network","cited_arxiv_id":null,"evidence_quote":"Provides RoI Aware Pooling used in the point-based branch to aggregate fine-grained local features."},{"cited_title":"Openpcdet: An open-source toolbox for 3d object detection from point clouds","cited_arxiv_id":null,"evidence_quote":"Supplies the open-source implementation base used for training and evaluation."}],"review_version":1}