{"id":"a3afe7cd-d39c-49b3-9eb7-7b93adf527ce","arxiv_id":"2507.10222","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Replicating 2D images into a 3D volume and processing with a shallow 3D U-Net gives competitive dense prediction accuracy at a fraction of the parameter count, with slice-consistency used as a free quality score.","lead":"A computer vision method called Spatial Lifting copies each 2D image into a stack of slices and processes the stack with a 3D U-Net, claiming competitive segmentation and depth results with far fewer parameters and lower compute. It also uses the consistency between slices to estimate prediction quality without extra cost.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (1) replicates the input exactly along z; a standard 3D U-Net is z-translation equivariant, so all interior output slices are identical, making slice selection and PQA vacuous.","rationale":"The reader's conditional pass rests on a slice-selection ablation, but the more fundamental question is whether slices can differ at all. Because Eq. (1) makes the input exactly constant along z and the standard 3D U-Net layers are translation-equivariant, the output slices should be identical in the interior; this follows from a simple symmetry/induction argument and is independent of training data or initialization. If true, it invalidates the paper's stated mechanism: slice selection is a tie-break among identical copies, PQA is measuring consistency with mostly itself, and the claimed 'intrinsically structured outputs' do not emerge. The empirical accuracy/efficiency numbers might still be reproducible, since summing identical slices is equivalent to one slice, but the core novelty and the PQA contribution would collapse. This is more load-bearing than the missing ablations because no amount of slice-selection ablation can fix a mechanism that cannot produce diversity. A concrete slice-identity check on the released code would settle the issue immediately; the paper currently provides no code, making the results unverifiable. I therefore recommend rejection pending that check, rather than conditional acceptance.","tokens_in":17830,"tokens_out":12030,"duration_ms":158613,"concrete_test":"Run the released or re-trained SL-UNet (8L, 2Res) on one dataset, e.g., ISIC2018. For a set of test images, compute the mean absolute difference and Dice between the output logits/masks at interior slices z=8 and z=9, and compare the final Dice obtained from the selected 5 slices against using only slice z=8. If the interior slices are identical to machine precision (diff ~ 0, Dice = 1) and the final Dice is unchanged, then Eq. (6)-(7) and Eq. (11) are vacuous and only a single effective 2D prediction is being evaluated.","verdict_should_be":"REJECT","load_bearing_attack":"The load-bearing issue sits upstream of the reader's slice-selection concern. Eq. (1) sets I'(x,y,z,c)=I(x,y,c) for every z, so the 3D input is fully invariant under translation along the new axis. The 3D U-Net used in Section IV-B is built from 3D convolutions, normalizations, nonlinearities, pooling, and upsampling, all equivariant under z-translation, and the paper describes no positional encoding, per-slice weights, or other symmetry-breaking mechanism. Therefore P(x,y,z,c) is constant over z for all interior slices; only O(L) boundary slices can differ, with L the number of layers. This makes Eq. (6)-(7) select among identical predictions and means the test result is equivalent to evaluating a single interior slice (or boundary artifacts). It also means Eq. (11) compares identical segmentation maps for most selected/unselected pairs, so the PQA correlations in Table IV cannot reflect the mechanism in Proposition III.4. Unless the released code contains an unspecified z-dependent operation, the claimed 'structured outputs along the lifted dimension' and the PQA contribution do not exist as described; the model is effectively a 2D convolution with kernels summed over z.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes Spatial Lifting (SL) for dense prediction. A 2D image is replicated m times along a new z-axis (Eq. 1) and processed by a 3D U-Net with a constant, small channel width (C*=8). All m output slices are supervised with the same 2D ground truth; at the end of training the s=5 slices with lowest average training loss are selected (Eqs. 6-7), and at test time the selected slice logits are summed (Eq. 9). The authors claim that this scheme outperforms or matches conventional 2D U-Nets with 32-512 channels on 13 semantic segmentation and 6 depth estimation datasets while reducing parameters by over 98% and lowering GMACs, that it produces structured outputs along z that enable a near-free prediction quality assessment (Eq. 11), and that it provides implicit regularization. Theoretical support is offered in Section III-C.","tokens_in":18065,"tokens_out":11247,"duration_ms":136531,"significance":"If the proposed mechanism were real, the paper would be significant: a simple dimensionality-lifting recipe that improves accuracy while cutting parameters by two orders of magnitude, plus a built-in uncertainty signal, would be of broad interest. The empirical evaluation is unusually broad (19 datasets, two tasks, multiple backbones/decoders), and the efficiency tables (Tables II and VI) are clearly presented. However, the central mechanism is called into question by the z-translation equivariance of the architecture, and the paper's own results do not resolve this; the PQA correlations in Table IV cannot be taken as evidence of the proposed mechanism without an explicit symmetry-breaking component or an ablation. The significance therefore hinges on a correction that is not presently in the manuscript.","major_comments":[{"comment":"The input I' is constant along the lifted z-axis, and the 3D U-Net used in Section IV-B is composed of 3D convolutions, ReLU, normalization, pooling, and upsampling, none of which break translation symmetry along z away from boundaries. The paper describes no positional encoding, per-slice embedding, or other symmetry-breaking mechanism. Consequently, for the bulk of the m interior slices, the output P(x,y,z,c) is identical, and the only z-variation arises from padding/pooling boundary effects. This makes the slice selection in Eqs. (6)-(7) and the PQA score in Eq. (11) measures of boundary artifacts rather than of the claimed 'structured outputs along the lifted dimension' (Section I, Proposition III.4). The authors must either introduce explicit per-slice position information (e.g., a learned z embedding) or demonstrate empirically that slice predictions differ in a way that reflects the proposed mechanism rather than boundary effects.","section":"III-A, Eq. (1)"},{"comment":"The selection of z1..zs as the s slices with lowest average training loss is a post-hoc selection on the training labels. The paper never ablates this choice against random s-slice subsets or against averaging all m slices. If the ranking of slices on training data does not transfer to test data, the numbers in Tables III, VII, and VIII could change substantially; and if, as suggested above, most slices are identical, the selection is vacuous. An ablation comparing training-loss-based selection, random fixed selection, and all-slice averaging is necessary before the test-time protocol can be trusted.","section":"III-A, Eqs. (6)-(7)"},{"comment":"The theoretical analysis does not currently support the empirical claims. Theorem III.1 establishes only a generic Lipschitz property of convolutions, which says nothing specific about the lifted input when that input is constant along z; Theorem III.2 and Proposition III.3 are heuristic analogies to Rademacher complexity and gradient averaging, not formal statements. More importantly, the claimed 'implicit regularization' and 'variance reduction' arguments presuppose that different slices compute different features, which is exactly what the z-translation equivariance issue denies. Proposition III.4, in particular, rests on the assumption of divergent predictions across z, and if most slices are identical the Q statistic cannot behave as described.","section":"III-C, Theorems III.1-III.2 and Propositions III.3-III.4"},{"comment":"The accuracy comparison is confounded by channel width. The 2D U-Net baselines use channel widths of 32-512, while SL-UNet uses 8 channels throughout. No matched-capacity 2D baseline (e.g., a 2D U-Net with 8 channels and comparable GMACs) is reported, so the observed Dice improvements cannot be attributed to dimensionality lifting per se. Since parameter/compute reduction is the paper's headline, a matched-width or matched-compute ablation is necessary to support the claim that spatial lifting, rather than some other training or architecture difference, is responsible for the results.","section":"IV-B, Tables II-III"}],"minor_comments":[{"comment":"For multi-class segmentation, the activation sigma should be a softmax over classes rather than a sigmoid; as written, summing logits and then applying a per-class sigmoid before argmax is not a standard probability combination for single-label segmentation.","section":"III-A, Eq. (9)"},{"comment":"The notation BM(.,.,z,.) is not defined before its use in the PQA score; please define the binarization procedure explicitly.","section":"III-B, Eq. (11)"},{"comment":"The text states that NYU Depth V2 has 795 training images, while Table I reports 50,688 training images; these numbers need to be reconciled.","section":"IV-A, Table I vs. text"},{"comment":"The text says the SL models achieve 'significant improvements' across most datasets, but several RMSE entries degrade (e.g., ResNeXt+SL-MiDaS on Make3D and KITTI); please quantify the trade-off and avoid the impression of uniform gains.","section":"IV-D, Table VIII"},{"comment":"The interpretability visualizations in Fig. 2 are not quantitatively connected to slice selection or to the PQA mechanism; consider showing which slices are selected and how their visualizations differ from unselected ones.","section":"II-D, Fig. 2"}],"recommendation":"reject","confidential_remarks":"The paper is well-organized and the empirical breadth is substantial, but the z-translation equivariance problem is fundamental: as written, the lifted dimension provides no meaningful structure beyond boundary artifacts. This is not a local presentation issue; the central claims about structured outputs, dense slice supervision, and prediction quality assessment all depend on a mechanism the architecture does not implement. A revision that adds explicit symmetry-breaking and reruns the experiments would be a different manuscript, and the slice-selection and matched-width ablations would need to be mandatory. I therefore recommend rejection of the current version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe stress-test note is right, and it sinks the central mechanism. Eq. (1) replicates the image identically along z, and a standard 3D U-Net is z-translation equivariant, so every output slice is identical (any boundary effects are washed out by downsampling to z=1). That makes the slice-selection step (Eq. 6-7) a selection among identical predictions, and the PQA mechanism (Eq. 11) a comparison between identical maps. The correlations in Table IV cannot be real under the architecture as described; either the code has an undocumented z-dependent operation or the results are from a different setup. This is a load-bearing flaw, not a cosmetic one.\n\nCredit where due: the paper is clearly written, the experiment set is wide (13 segmentation + 6 depth), and the parameter reduction is genuine—SL-UNet is about 100x smaller. The simple idea of lifting a 2D image to 3D and using a low-channel 3D U-Net is worth a look, though prior pseudo-3D and 2.5D work is not cited. The theoretical sections are proof sketches and add little.\n\nThe soft spots beyond the symmetry issue: the baselines are not parameter-matched. SL-UNet uses 8 channels throughout; the 2D U-Nets use 32-512. A fair comparison would be a 2D U-Net with 8 channels, and the reported gains may mostly be channel reduction rather than lifting. The slice selection is never ablated against random or all-slice averaging. And the PQA claim is vacuous if the slices are identical.\n\nWho would get value from this? Someone interested in a cautionary tale about symmetry assumptions, or in lightweight decoders if the authors can actually add a z-dependent operation and re-run the experiments. As written, the paper's two headline contributions—structured outputs and intrinsic quality assessment—do not exist. It deserves a desk reject, not a conditional pass. If the authors revise with a real z-dependence and fair baselines, it could become a modest efficiency paper.\n\nRecommendation: desk reject in current form.","headline":"The z-symmetry argument holds: the replicated input makes output slices identical, so the paper's slice-selection and PQA contributions are vacuous; the empirical results are likely real but the framing is wrong.","tokens_in":18635,"tokens_out":6831,"would_cite":false,"duration_ms":79890,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Spatial Lifting claims that copying a 2D image into a third dimension and processing it with a channel-constant 3D U-Net can match or beat conventional 2D dense prediction networks while cutting parameters by over 98%.","keywords":["spatial lifting","dense prediction","semantic segmentation","depth estimation","3D U-Net","prediction quality assessment","model parameter reduction","implicit regularization"],"falsifier":"On a held-out dataset, compute the per-slice Dice of each of the $m=16$ output slices at test time and compare the ranking with the training-loss ranking used to select the five slices; if the selected slices are not among the best test slices, or if averaging all 16 slices beats the selected five, the slice-selection step is not doing the claimed work.","tokens_in":17573,"feed_emoji":"🧊","tokens_out":8376,"duration_ms":82403,"temperature":0.7,"pith_summary":"This paper proposes Spatial Lifting (SL), a method that takes a standard 2D image, copies it sixteen times along a new third axis, and feeds the resulting 3D volume to a 3D U-Net that keeps only 8 channels at every depth. The paper's central claim is that this counterintuitive dimensionality increase lets the small 3D model match or beat a conventional 2D U-Net with channels growing from 32 to 512, on 13 semantic segmentation datasets and 6 depth estimation datasets. If the claim holds, dense prediction models could shrink by over 98% in parameters and lower inference cost while staying competitive, and the replicated output slices provide a near-free per-image quality score. A sympathetic reading is that SL is offered as a new modeling paradigm: trade channel width for an extra spatial axis and get efficiency plus built-in uncertainty estimation.","feed_headline":"Lifting images to 3D cuts U-Net parameters by 98%","feed_subtitle":"A 3D U-Net with just 8 channels matches or beats 2D U-Nets on 19 datasets at a fraction of the cost.","key_machinery":"The central object is the lifted volume: an input image replicated $m=16$ times along a new depth axis, producing a 3D tensor that a 3D U-Net processes into $m$ output slice predictions. The key identity is the complexity trade-off: a 2D U-Net's parameters grow like the product of successive channel counts (32×64×...×512), while SL keeps a constant $C_*=8$ channels so parameters scale with $C_*^2$ times the kernel volume, and MACs only pick up a factor $m \\cdot K_z$. The third mechanism is slice selection and quality scoring: after training, the $s=5$ slices with the lowest average training loss (Eq. 6-7) are selected and their logits are summed at test time, and the average Dice between selected and unselected slices defines the quality score $Q$.","core_discovery":"On its own terms, the paper discovers that spatial lifting is a viable alternative to channel-widening in dense prediction. A 3D U-Net with constant 8 channels, fed with a 16-slice lifted volume, is reported to match or outperform a 2D U-Net with 32-512 channels on most of the 13 segmentation datasets (e.g., Dice 90.00 vs 87.76 on DSB2018 for the 5L,2Res config) and to improve depth RMSE and δ1 on most of the 6 depth datasets, with over 98% parameter reduction and lower GMACs in the U-Net case. The paper also claims that the lifted output's slice structure enables a prediction-quality score Q—the mean Dice agreement between the five selected slices and the remaining slices—that correlates with true Dice (Pearson r up to 0.92 on FIVES, mostly significant). The theoretical contribution is the argument that 3D convolutions with shared weights along the lifted axis impose an implicit Lipschitz constraint, giving a smoothness bias and improved generalization, while dense supervision across slices acts as a gradient-averaging regularizer.","pith_inferences":["The slice-selection step is the least guarded part of the pipeline: if the training-loss ranking of slices does not transfer to test data, an adaptive test-time selection rule (for example, picking the most mutually consistent slices) could be tried instead.","The lift factor $m$ is fixed at 16 in all experiments; a scaling study varying $m$ would clarify whether the gains come from having many slices per se or from the channel reduction, and would give practical guidance for new tasks.","The same recipe might transfer to other input dimensions, such as lifting video or point clouds into a 4D volume, where the extra axis could act as a pseudo-time or pseudo-scale dimension.","Since $Q$ measures cross-slice agreement, it could be turned into a training objective that explicitly enforces slice consistency, which the paper does not currently propose."],"forward_implications":["Dense prediction models can get accurate results with tens of thousands of parameters instead of millions, by replacing channel depth with an extra spatial axis.","Every SL model outputs a per-image quality estimate at negligible extra cost, which could be used to flag low-confidence predictions in deployment.","SL transfers across backbones and decoders: the paper reports gains or parity when applied to HSNet, CASCADE, EMCAD, FastDepth, and MiDaS on most datasets, so it functions as a plug-in design principle, not just a single architecture.","The implicit-regularization story predicts that SL models should degrade smoothly under perturbation or distribution shift along the lifted axis, because the shared 3D weights enforce z-smoothness."],"supporting_citations":[{"why":"Supplies the base 2D U-Net architecture whose channel-widening design is the main baseline in Tables II and III.","marker":"[1]"},{"why":"Provides the 3D U-Net used as the higher-dimensional processing network in SL.","marker":"[17]"},{"why":"Supplies the pretrained PVTv2 encoder used in the decoder experiments with SL.","marker":"[30]"},{"why":"Supplies the MiDaS depth decoder that SL is integrated into and compared against in Table VIII.","marker":"[3]"},{"why":"Supplies the FastDepth decoder baseline for depth estimation.","marker":"[60]"},{"why":"Supplies the EMCAD efficient decoder baseline and its SL-enhanced variant.","marker":"[56]"},{"why":"Supplies the HSNet decoder baseline in the decoder comparison.","marker":"[55]"},{"why":"Represents MC dropout, the standard uncertainty method that SL's near-zero-cost prediction quality assessment is contrasted with.","marker":"[10]"}],"fun_headline_variants":["Spatial Lifting: 3D U-Net beats 2D, cuts params 98%","New paradigm: spatial lifting cuts U-Net params by 98%","Lifting to 3D: 98% fewer params, better dense prediction","3D U-Net with 8 channels matches 2D, uses 98% fewer params","Spatial Lifting: new dense prediction paradigm, 98% leaner"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the five output slices with the lowest average training loss are also the best slices to average at test time; the paper does not ablate this selection against using all $m$ slices or a fixed subset.","fun_headline_variants_meta":{"raw":{"variants":["Spatial Lifting: 3D U-Net beats 2D, cuts params 98%","New paradigm: spatial lifting cuts U-Net params by 98%","Lifting to 3D: 98% fewer params, better dense prediction","3D U-Net with 8 channels matches 2D, uses 98% fewer params","Spatial Lifting: new dense prediction paradigm, 98% leaner"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000715,"raw_usage":{"total_tokens":3228,"prompt_tokens":974,"completion_tokens":2254,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":2140}},"tokens_in":590,"tokens_out":2254,"duration_ms":15083,"temperature":1.0,"reasoning_tokens":2140,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:36:15.269683+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a held-out dataset, compute the per-slice Dice of each of the $m=16$ output slices at test time and compare the ranking with the training-loss ranking used to select the five slices; if the selected slices are not among the best test slices, or if averaging all 16 slices beats the selected five, the slice-selection step is not doing the claimed work.","supporting_citations":[{"cited_title":"U-net: Convolutional networks for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the base 2D U-Net architecture whose channel-widening design is the main baseline in Tables II and III."},{"cited_title":"3d u-net: learning dense volumetric segmentation from sparse annotation,","cited_arxiv_id":null,"evidence_quote":"Provides the 3D U-Net used as the higher-dimensional processing network in SL."},{"cited_title":"Local rademacher complexities,","cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained PVTv2 encoder used in the decoder experiments with SL."},{"cited_title":"Towards robust monocular depth estimation: Mixing datasets for zero-shot cross- dataset transfer,","cited_arxiv_id":null,"evidence_quote":"Supplies the MiDaS depth decoder that SL is integrated into and compared against in Table VIII."},{"cited_title":"Aggregated residual transformations for deep neural networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the FastDepth decoder baseline for depth estimation."},{"cited_title":"Hsnet: A hybrid semantic network for polyp segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the EMCAD efficient decoder baseline and its SL-enhanced variant."},{"cited_title":"Modest museum dataset,","cited_arxiv_id":null,"evidence_quote":"Supplies the HSNet decoder baseline in the decoder comparison."},{"cited_title":"Dropout as a bayesian approximation: Representing model uncertainty in deep learning,","cited_arxiv_id":null,"evidence_quote":"Represents MC dropout, the standard uncertainty method that SL's near-zero-cost prediction quality assessment is contrasted with."}],"review_version":1}