{"id":"e7ba7d88-e850-465f-982e-d35a11985484","arxiv_id":"2502.07289","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"LP-Net reconstructs dense depth maps by progressively refining a low-resolution estimate with learned smoothness and sharpness filters, reporting state-of-the-art KITTI accuracy at lower compute.","lead":"LP-Net is a depth completion network that first predicts a low-resolution, blurry depth map and then sharpens it incrementally with learned filters. It reports top KITTI leaderboard accuracy while running faster and using less memory than recent propagation-based methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SDF sharpness filter as written sums its kernel to zero, so it outputs zero on flat regions; Eq. 8-9 cannot produce the reported sharp boundaries.","rationale":"The paper's central claim is a state-of-the-art accuracy-efficiency trade-off via a multi-scale Laplacian-pyramid scheme and a single-pass Selective Depth Filtering module. The most load-bearing point is the internal consistency of that module. The sharpness filter as written applies a zero-sum convolution, which is a high-pass/derivative operator, not a sharpening operator. Substituting it into Eq. 9 produces a final depth that collapses toward zero wherever the attention map selects the sharpness branch. This is a concrete mathematical problem, not merely a missing ablation: it means the module cannot behave as described, and the reported qualitative sharpening at boundaries is unexplained. The reported KITTI leaderboard numbers are external evidence that some implementation works, so the issue may be a typo in the paper; but as submitted, the method is not reproducible from the text. The reader's concern about single-pass sufficiency is reasonable and related, but the sum-zero normalization is more fundamental: if the sharpness branch is mis-specified, iteration count is secondary. The NYUv2 claims are also overstated in the text (Table II lists several methods with RMSE 0.089 versus LP-Net's 0.090), and no code is released, both noted by the reader. These issues together support a conditional verdict: the paper should not be accepted without either a corrected and consistent SDF formulation or the released code confirming the reported numbers. Since the reader already recommended CONDITIONAL, my stress test does not change that verdict.","tokens_in":19538,"tokens_out":11051,"duration_ms":104399,"concrete_test":"Implement the SDF module exactly as written in Eqs. 7-9: use deformable convolution with per-pixel weights normalized to zero sum for fa, feed a constant depth patch, and evaluate Eq. 9 with attention a=0. The output is identically zero, which falsifies the described sharpening behavior. If the authors release code, inspect the sharpness branch: if the released kernel sums to 1 or includes a residual addition, the paper's text is wrong; otherwise the reported KITTI validation RMSE of 709.82 mm cannot be reproduced by the described module.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section III-C3, the paper specifies that the sharpness filter weights are tanh-constrained and then normalized by subtracting their mean, 'ensuring they sum to 0' (Eq. 8). A convolution with zero-sum weights maps any constant input to zero. Since D'_m is the output of the smoothness filter and is approximately constant over smooth regions, D'_a is approximately zero there. Eq. 9 then outputs Dhat(i) = a*D'_m + (1-a)*D'_a. At boundaries, where the selection map in Fig. 7 is said to favor the sharpness branch (1-a high), this formula drives the final depth toward zero instead of sharpening the boundary. No residual connection adds D'_m back to D'_a, and no sum-to-one normalization is described. The reported sharp boundary recovery in Figs. 8-10 therefore cannot be produced by the module exactly as written. Either Eq. 8 should sum to one, or Eq. 9 should include a residual term, or the reported experiments use a different implementation. Because SDF is the paper's key novelty and the basis for the single-pass efficiency claim, this is a load-bearing internal inconsistency.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LP-Net, a depth completion network built around an inverse Laplacian pyramid: it first regresses a low-resolution depth residual, then progressively upsamples and refines the depth through four additional scales. Two modules are introduced: the Multi-path Feature Pyramid (MFP), which splits features into parallel pathways with different downsampling factors, and the Selective Depth Filtering (SDF) module, which learns a smoothness filter and a zero-sum high-pass sharpness filter, combined by a learned spatial selection map. Experiments report state-of-the-art results on KITTI DC using official leaderboard numbers, competitive results on NYUv2, and best results on TOFDC, alongside lower inference time and memory than recent propagation-based methods.","tokens_in":19866,"tokens_out":7477,"duration_ms":65392,"significance":"If the results are reproducible, LP-Net is a meaningful contribution: it challenges the prevailing iterative spatial-propagation paradigm, offers a strong accuracy-efficiency trade-off, and introduces a progressive multi-scale architecture for depth completion. The use of externally hosted KITTI leaderboard numbers and standard public benchmarks is a credible evaluation strategy. The main limitations are the mathematical description of the key SDF module, which appears internally inconsistent, and an overstatement of the NYUv2 rankings. These issues affect the paper's central claims and need to be resolved before the work can be fully accepted.","major_comments":[{"comment":"The sharpness filter weights are explicitly normalized to sum to zero by subtracting the mean (Eq. 8), so the output D'_a is a high-pass residual: it is approximately zero on smooth regions and, by construction, cannot carry absolute depth information. Eq. (9) then forms a convex combination D_hat = a*D'_m + (1-a)*D'_a. Since the text states that the selection map favors the sharpness branch (1-a high) at boundaries, the final output at boundaries would be dominated by a zero-mean high-pass signal, which cannot plausibly reconstruct the surface depth there. This is also inconsistent with the Laplacian pyramid reconstruction in Eq. (2), which is additive. Please either change Eq. (9) to an additive residual form (e.g., D_hat = D'_m + (1-a)*D'_a), or normalize the sharpness kernel to sum to one, or provide a concrete explanation and empirical evidence that the current convex combination produces valid depth values in the sharpness-selected regions.","section":"Section III-C3 (Eqs. 8-9)"},{"comment":"The text claims LP-Net achieves 'second-best performance in RMSE and REL' on NYUv2. This is not supported by Table II: BP-Net, DySPN, BEV@DC, and PointDC all report RMSE 0.089 while LP-Net reports 0.090; for REL, LRRU and DFU report 0.011 while LP-Net reports 0.012. The claim overstates the method's ranking and should be corrected to match the table, or the table must be revised.","section":"Section IV-D2 (Table II)"}],"minor_comments":[{"comment":"In the RMSE column of Variant-iii, '721,83' uses a comma as the decimal separator; it should be '721.83'.","section":"Table VI"},{"comment":"The paper's SDF module is described as a key component of the inverse Laplacian pyramid, but Eq. (9) is a weighted average rather than an additive reconstruction. The figure caption mentions an 'Addition' operation; please align the figure with the equation or clarify the relationship to Eq. (2).","section":"Section III-B and Fig. 6"},{"comment":"If the sharpness filter is intended as a high-pass residual (unsharp-mask style), please cite the relevant image-enhancement literature and clearly define D'_a as a residual so that the formulation in Eq. (9) can be interpreted correctly.","section":"Section III-C3"},{"comment":"The ablation tables report single-run results without error bars. Given that several metric differences (e.g., MAE 185.38 vs 186.10 in Table V) are small, reporting mean and standard deviation over multiple runs would strengthen the component-wise conclusions.","section":"Section IV-E"}],"recommendation":"major_revision","confidential_remarks":"The paper's connection to the authors' prior work, especially TPVD (CVPR 2024) and DCDepth (NeurIPS 2024), should be clarified; the progressive multi-scale idea resembles DCDepth's DCT-domain approach, and the novelty relative to those works should be explicitly discussed. The claim of ranking 1st on the KITTI leaderboard is time-dependent and should be accompanied by the submission date. Public code would substantially aid verification of the SDF equations and the reported efficiency numbers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a plausible efficiency-first depth completion design, and the KITTI leaderboard results are real evidence. But the SDF module has a load-bearing flaw in the math that the authors need to resolve.\n\nWhat is new: an inverse Laplacian pyramid progressive prediction scheme, a Multi-path Feature Pyramid module, and a Selective Depth Filtering module that applies learned smoothness and sharpness filters once rather than iteratively. The KITTI numbers are externally hosted, speed and memory are measured on a single RTX 4090, and the ablations show MFP and SDF each contribute. That is solid, reproducible-style evidence for an engineering paper.\n\nThe soft spot is serious. Section III-C3 says the sharpness filter weights are tanh-constrained then normalized by subtracting the mean so they sum to zero. A zero-sum convolution maps constant input to zero, so after the smoothness filter the sharpness branch output D'_a is near zero on flat regions. Equation (9) then computes a convex combination a*D'_m + (1-a)*D'_a, and at boundaries where the attention map favors the sharpness branch, the output is driven toward a high-pass response rather than the actual depth value. The sharp boundaries shown in Figs. 8-10 cannot come from this formula. What was probably intended is a residual formulation like D_hat = D'_m + (1-a)*D'_a, or a sharpness kernel that preserves the local mean. As written, the key module is internally inconsistent, and the single-pass efficiency claim rests on it.\n\nOther issues are smaller: the NYUv2 section calls the performance SOTA, but Table II shows LP-Net is second-best in RMSE and REL; the KITTI RMSE gain over BP-Net is 0.19 mm, which is inside the noise; there is no code, no error bars, and Table VI has a typo (721,83). Those are addressable.\n\nWho this is for: people working on efficient depth completion, especially for autonomous driving. The paper deserves a serious referee, but mainly so the authors can be pushed to release code or correct the SDF formulation. I would not accept it as-is.","headline":"LP-Net has a real efficiency story and strong external KITTI numbers, but the Selective Depth Filtering module as written is internally inconsistent and needs fixing before the results can be trusted.","tokens_in":20312,"tokens_out":3315,"would_cite":false,"duration_ms":34557,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"LP-Net runs the Laplacian pyramid in reverse: it predicts a coarse global depth, then progressively re-adds high-frequency detail via a learned smoothness-versus-sharpness filter, and reports top KITTI accuracy with each filter applied…","keywords":["depth completion","Laplacian pyramid","multi-scale prediction","selective depth filtering","sparse depth","RGB-guided depth completion","progressive refinement","feature pyramid"],"falsifier":"Run LP-Net inference with the smoothness and sharpness filters applied a second time at each scale on KITTI validation and compare RMSE and MAE to the single-pass model; if the second pass improves RMSE by more than a small margin, the claim that a single non-iterative filtering pass recovers the high-frequency detail is falsified.","tokens_in":19389,"feed_emoji":"🗺️","tokens_out":6317,"duration_ms":55038,"temperature":0.7,"pith_summary":"LP-Net argues that depth completion should not be formulated as iterative propagation from a coarse estimate, but as the inverse of a Laplacian pyramid: first predict the low-frequency residual at 1/16 resolution to capture the whole scene, then upsample and restore the missing high-frequency structure one scale at a time. The paper introduces two modules, a Multi-path Feature Pyramid for global context and a Selective Depth Filtering module that learns both a smoothness and a sharpness filter, and shows that this progressive scheme reaches the top of the KITTI Depth Completion leaderboard and the best RMSE on TOFDC while running faster and using less memory than recent propagation-based competitors. The results matter because the dominant family of single-scale propagation methods pays a heavy computational cost for many iterations of pixel-level message passing; if the pyramid formulation is right, accurate dense depth can be obtained with a single forward pass of cheap per-scale filtering.","feed_headline":"Inverse Laplacian pyramid tops KITTI depth, 23% faster","feed_subtitle":"New depth completion network beats recent propagation rivals on accuracy while using 1.76 GB memory and 63.88 ms per frame.","key_machinery":"The load-bearing object is the Laplacian pyramid, used in reverse. Instead of decomposing an image into bandpass levels and a low-frequency residual, LP-Net predicts the low-frequency residual first and then reconstructs the pyramid upward. The Multi-path Feature Pyramid (MFP) module splits the deepest feature map into p pathways, applies different numbers of stride-2 convolutions per pathway, and re-fuses them to capture global context cheaply. The Selective Depth Filtering (SDF) module is the refinement engine: it generates deformable filter weights and offsets, builds a smoothness filter whose weights are softmax-normalized to sum to 1 and a sharpness filter whose tanh-normalized, mean-subtracted weights sum to 0, applies each exactly once, and blends the two outputs with a learned spatial attention map. A confidence mechanism fuses each predicted scale with the weighted pooling of the sparse input, and a multi-scale L1 plus L2 loss supervises every level.","core_discovery":"The paper's central claim is that a dense depth map can be generated by mirroring Laplacian pyramid reconstruction. The network starts from the lowest-resolution prediction, fuses it with a confidence-weighted downsampled sparse depth, and at each higher scale bilinearly upsamples the previous estimate, fuses the sparse measurements, and applies a learned selection between a smoothing filter whose kernel weights sum to 1 and a sharpening filter whose kernel weights sum to 0, both with deformable offsets. This replaces the iterative recurrent propagation of CSPN-style methods with a strictly progressive, once-per-scale refinement. On the official KITTI DC test set LP-Net reaches RMSE 684.71 mm, MAE 186.63 mm, iRMSE 1.81, and iMAE 0.80, and it reports 63.88 ms inference and 1.76 GB GPU memory on an RTX 4090, which the paper attributes precisely to avoiding iteration.","pith_inferences":["Editorial inference: the inverse-pyramid formulation is mathematically equivalent to learning a set of bandpass residuals, so the intermediate predictions D(4) through D(1) could be treated as explicit residual channels and inspected or regularized directly, a diagnostic the paper does not run.","Editorial inference: because the filters are applied once, the receptive field of the full-resolution stage is small; very thin or very distant structures that require global reasoning may be the first failure mode, which is testable by stress-testing on long thin objects.","Editorial inference: the same progressive completion schedule could be applied to other sparse-to-dense tasks such as depth super-resolution, normal completion, or sparse-view stereo, where the bottleneck is the recovery of high-frequency detail rather than coarse layout."],"forward_implications":["If the claims hold, iterative spatial propagation is not necessary for top-tier depth completion, so the field can shift from recurrent refinement to feed-forward multi-scale refinement.","The progressive steps can be truncated: earlier pyramid levels already produce usable depth, so the same network can trade accuracy for latency at runtime by stopping after fewer upsampling stages.","The success of MFP over a ViT block in ablations suggests global scene context in depth completion can be gathered with simple multi-scale convolutions instead of self-attention, reducing cost.","A learned per-pixel blend between a smoothing and a sharpening filter is enough to denoise interiors while preserving boundaries, a recipe likely to transfer to other image-guided dense prediction problems.","Demonstrated on a LiDAR outdoor benchmark, a Kinect indoor benchmark, and a ToF indoor benchmark, the pyramid scheme is not tied to one sensor type or depth range."],"supporting_citations":[{"why":"Defines the convolutional spatial propagation baseline whose iterative message passing LP-Net replaces and the main efficiency contrast.","marker":"[15]"},{"why":"Establishes the non-local propagation baseline and the NYUv2 evaluation protocol LP-Net follows.","marker":"[16]"},{"why":"Supplies the weighted pooling confidence fusion technique and represents BP-Net, the closest KITTI competitor.","marker":"[30]"},{"why":"Introduces the TOFDC dataset and the tri-perspective view competitor on which LP-Net reports the best RMSE.","marker":"[36]"},{"why":"LRRU is a recent recurrent propagation method used as a sharpness-related baseline in the efficiency comparison.","marker":"[39]"},{"why":"DFU is a propagation-based competitor with a depth feature upsampling design that LP-Net compares against.","marker":"[42]"},{"why":"The U-Net architecture provides the encoder-decoder structure that LP-Net builds on.","marker":"[40]"},{"why":"Deformable convolution supplies the offset mechanism used by the adaptive kernels in the SDF module.","marker":"[43]"},{"why":"The KITTI depth completion benchmark is the source of the official leaderboard numbers.","marker":"[54]"},{"why":"The NYUv2 dataset is used for the indoor evaluation and for the 500-point sparse depth protocol.","marker":"[55]"}],"fun_headline_variants":["Laplacian pyramid depth completion tops KITTI, cuts compute","Progressive pyramid refinement wins depth benchmarks at lower cost","LP-Net: multi-scale depth from sparse input, SOTA on KITTI","Single-pass depth pyramid beats iterative rivals on KITTI","Inverse pyramid depth net: faster and more accurate on KITTI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole speed and memory advantage rests on one premise: a single pass of the learned smoothness and sharpness filters at each resolution is enough to restore the missing detail; if those filters must be iterated many times to reach the reported accuracy, the efficiency claim disappears.","fun_headline_variants_meta":{"raw":{"variants":["Laplacian pyramid depth completion tops KITTI, cuts compute","Progressive pyramid refinement wins depth benchmarks at lower cost","LP-Net: multi-scale depth from sparse input, SOTA on KITTI","Single-pass depth pyramid beats iterative rivals on KITTI","Inverse pyramid depth net: faster and more accurate on KITTI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000255,"raw_usage":{"total_tokens":1610,"prompt_tokens":1022,"completion_tokens":588,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":638,"completion_tokens_details":{"reasoning_tokens":497}},"tokens_in":638,"tokens_out":588,"duration_ms":6029,"temperature":1.0,"reasoning_tokens":497,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T13:11:55.937699+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run LP-Net inference with the smoothness and sharpness filters applied a second time at each scale on KITTI validation and compare RMSE and MAE to the single-pass model; if the second pass improves RMSE by more than a small margin, the claim that a single non-iterative filtering pass recovers the high-frequency detail is falsified.","supporting_citations":[{"cited_title":"Learning depth with convolutional spatial propagation network,","cited_arxiv_id":null,"evidence_quote":"Defines the convolutional spatial propagation baseline whose iterative message passing LP-Net replaces and the main efficiency contrast."},{"cited_title":"Non-local spatial propagation network for depth completion,","cited_arxiv_id":null,"evidence_quote":"Establishes the non-local propagation baseline and the NYUv2 evaluation protocol LP-Net follows."},{"cited_title":"Bilateral propagation network for depth completion,","cited_arxiv_id":null,"evidence_quote":"Supplies the weighted pooling confidence fusion technique and represents BP-Net, the closest KITTI competitor."},{"cited_title":"Tri-perspective view decomposition for geometry-aware depth completion,","cited_arxiv_id":null,"evidence_quote":"Introduces the TOFDC dataset and the tri-perspective view competitor on which LP-Net reports the best RMSE."},{"cited_title":"Lrru: Long- short range recurrent updating networks for depth completion,","cited_arxiv_id":null,"evidence_quote":"LRRU is a recent recurrent propagation method used as a sharpness-related baseline in the efficiency comparison."},{"cited_title":"Improving depth completion via depth feature upsampling,","cited_arxiv_id":null,"evidence_quote":"DFU is a propagation-based competitor with a depth feature upsampling design that LP-Net compares against."},{"cited_title":"Deformable convolutional networks,","cited_arxiv_id":null,"evidence_quote":"Deformable convolution supplies the offset mechanism used by the adaptive kernels in the SDF module."},{"cited_title":"Are we ready for autonomous driving? the kitti vision benchmark suite,","cited_arxiv_id":null,"evidence_quote":"The KITTI depth completion benchmark is the source of the official leaderboard numbers."},{"cited_title":"Indoor segmentation and support inference from rgbd images,","cited_arxiv_id":null,"evidence_quote":"The NYUv2 dataset is used for the indoor evaluation and for the 500-point sparse depth protocol."}],"review_version":1}