{"id":"84b25cd8-3dfe-44c1-9ed8-18e514f7d783","arxiv_id":"2505.23093","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"LeMoRe combines Cartesian axis-wise views with nested attention to reach 33.5% mIoU on ADE20K at 0.8 GFLOPs and 1.6M parameters, a favorable efficiency point but not an accuracy leader.","lead":"LeMoRe is a small neural network for semantic segmentation that combines cheap axis-wise feature views with a compact nested attention mechanism. It reports lower compute and parameter counts than many lightweight baselines, though on most benchmarks its accuracy is below the strongest models.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline efficiency numbers in Tables 1, 3, 4, and 5 are not measured under a common input-resolution/hardware protocol; the paper's own LeMoRe* GFLOPs row contradicts its 512x512 footnote, so the 'balanced trade-off' claim rests on unverified comparisons.","rationale":"I agree with the reader that the weakest point is the comparability of efficiency numbers, and I found corroborating evidence inside the paper: the LeMoRe* row's 0.6 GFLOPs is numerically consistent with 448x448 inference, not the stated 512x512 protocol, and the Cityscapes table mixes FLOPs computed at different resolutions across rows. Since the paper's headline is explicitly an efficiency-accuracy trade-off, these inconsistencies are not cosmetic. The reader's CONDITIONAL verdict is the right one: the architecture is plausible and the ablation table is internally consistent, but the quantitative foundation cannot be checked without code and a unified benchmark. I do not see an internal contradiction in the ablation numbers or a mathematical error in the attention equations serious enough to overturn the main claim; the issue is the evidential basis of the comparison. Therefore I leave the verdict unchanged and propose a re-measurement test that would settle whether the claimed margin survives.","tokens_in":10234,"tokens_out":13670,"duration_ms":118887,"concrete_test":"Use the released checkpoint/code and a standard FLOP counter plus a fixed GPU to re-measure LeMoRe and every baseline in Tables 1, 3, 4, and 5 at the exact input resolutions used for that dataset (512x512 for ADE20K/COCO-Stuff, 1024x512 for Cityscapes, 480x480 for PASCAL Context), in the same framework and batch size; also measure latency on the same GPU. If LR-ASPP-MobileNetV3-Large comes out at 1.3 GFLOPs instead of 2.0, or if LeMoRe*'s 512x512 GFLOPs is 0.8 not 0.6, the table-level efficiency advantage is not as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is a trade-off statement: 33.5 mIoU at 0.8 GFLOPs, 1.6M params, and 24 ms on ADE20K. Its persuasive force comes from the baseline rows in Tables 1, 3, 4, and 5, but those numbers do not share a stated measurement protocol. Table 1 says GFLOPs are for 512x512 input, yet the LeMoRe* row (0.6 GFLOPs, 23 ms) is marked 'trained with 448x448' and 0.8*(448/512)^2 ~ 0.61, indicating its FLOPs were actually computed at 448; if they were recomputed at 512, the row would equal ~0.8 GFLOPs, contradicting the table. Tables 3-5 have no resolution note: Cityscapes LeMoRe at 1.7 GFLOPs matches a 1024x512 input (about 2x the 512x512 ADE value), while the PSPNet baseline is 423.4 GFLOPs, a number that only makes sense at roughly 2048x1024, not at 1024x512; LR-ASPP's 9.7 GFLOPs lies between the two. Latency is absent for most baselines, and no hardware is specified. Because the claimed reductions ('60% fewer params', '86.9% GFLOPs reduction', '38.5% cost reduction') mix resolutions and sources, a fair re-measurement could substantially shrink the efficiency gap. This is an unmet burden of quantitative comparability for a paper whose entire contribution is the efficiency/accuracy balance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces LeMoRe, a lightweight semantic segmentation architecture that combines explicit Cartesian projections (three 1x1 convolutions applied to spatially permuted feature maps) with implicit views learned through a nested attention mechanism, followed by a gated fusion decoder. The authors report 33.5 mIoU at 0.8 GFLOPs, 1.6M parameters, and 24 ms latency on the ADE20K validation set, together with results on Cityscapes, PASCAL Context, COCO-Stuff, and an object-detection experiment. The central claim is that LeMoRe achieves a superior accuracy/efficiency trade-off compared with lightweight transformer- and CNN-based baselines.","tokens_in":10522,"tokens_out":8458,"duration_ms":82123,"significance":"If the reported trade-off can be confirmed under a common measurement protocol, LeMoRe is a useful contribution to lightweight semantic segmentation: the architecture is simple, transparent, and considerably cheaper than transformer-based lightweight baselines, and the manuscript includes code, per-module ablations, and standard datasets/tooling that aid reproducibility. However, the quantitative basis of the central efficiency claim is currently weakened by inconsistent input-resolution reporting across tables, missing latency/hardware details for most baselines, and several unproven architectural assertions. The paper is therefore a promising candidate for publication after a revision that makes the efficiency comparisons like-for-like.","major_comments":[{"comment":"The GFLOPs reporting protocol is internally inconsistent. The table footnote states that GFLOPs are reported for 512x512 input, but the LeMoRe* row, marked as trained with 448x448 input, reports 0.6 GFLOPs; this is exactly 0.8*(448/512)^2, so the number is evidently computed at 448 resolution and would become 0.8 GFLOPs if recomputed at 512, making it identical to the LeMoRe row. Please recompute all FLOPs at one common resolution and clarify what distinguishes LeMoRe* from LeMoRe in cost terms.","section":"Section 4.1, Table 1"},{"comment":"The efficiency numbers in Tables 3-5 mix input resolutions without disclosure. The supplementary states that Cityscapes training uses a 1024x512 crop, so LeMoRe's 1.7 GFLOPs is about twice the 512x512 ADE20K figure; the PSPNet baseline of 423.4 GFLOPs is not plausible at 1024x512 (its ADE20K value at 512x512 is 52.2, implying a roughly 2048x1024 input). The claimed '99.60% reduction' in Section 4.2 is therefore not a like-for-like comparison. Specify the input resolution and source of every FLOPs number in Tables 3-5, or re-measure all baselines under one protocol, and state the hardware and measurement conditions for any latency values.","section":"Tables 3-5, Section 4.2"},{"comment":"The comparison to LR-ASPP is numerically inconsistent. Against LR-ASPP (MobileNetV3-Large) in Table 1, LeMoRe reduces GFLOPs by 60% (1 - 0.8/2.0) and parameters by 50%, not by 38.5%; the 38.5% figure matches a comparison with MobileNetV3-Large-reduce (1 - 0.8/1.3), for which the mIoU difference is +1.2, not +1.0. Please recompute all relative percentages and explicitly identify the baseline used for each claim.","section":"Section 4.1, last paragraph"},{"comment":"The claims that the Cartesian views form an 'orthogonal basis' and 'maximize joint entropy' are unsupported. A spatial permutation followed by a 1x1 convolution does not by itself enforce orthogonality among the resulting views, and no entropy or mutual-information computation is provided. Since these properties are invoked as the motivation for the decomposition, please supply a derivation or remove the claims.","section":"Section 3.1, Eq. (1)"},{"comment":"The Nested Attention mechanism is not fully specified. Eq. (3) aggregates pairwise products Qi⊗Kj for i,j=1,2,3 without softmax, scaling, or any normalization, so the resulting matrix is not a valid attention weighting; Eq. (4) then uses this matrix to weight value vectors. Please define the exact normalization and report the computational complexity of the nine pairwise interactions, since the efficiency argument for Nested Attention depends on this.","section":"Section 3.2, Eqs. (2)-(4)"},{"comment":"No repeated runs or error bars are reported. Several central comparisons are small (33.5 vs 33.1 for LR-ASPP; 32.2 vs 32.0 for R-ASPP; ablation increments of 0.4-0.5 mIoU), so run-to-run variation could change the conclusions. Please report mean±std over at least three seeds for the main models and ablations, or provide an equivalent significance assessment.","section":"Tables 1-5"}],"minor_comments":[{"comment":"There are minor language errors: 'comprises of' should be 'comprises' and 'In nutshell' should be 'In a nutshell'.","section":"Abstract and Section 4.1"},{"comment":"The symbol ¯⊗ is used in the Gated Fusion formula but is never defined; please define all operators in the equation.","section":"Eq. (5)"},{"comment":"The object-detection experiment lacks a dataset, input resolution, training schedule, and an explanation of how LeMoRe is adapted to RetinaNet; please add full experimental details or remove this section.","section":"Section 4.3, Table 6"},{"comment":"Reference [1] is a human-factors study about automated driving, which appears unrelated to the claim being cited; please replace it with an appropriate computer-vision reference.","section":"Introduction, reference [1]"},{"comment":"The latency values and their units (ms) are given in the table and plot, but no hardware or software stack is specified; please state the measurement conditions for all latency numbers.","section":"Figure 1(b) and Table 1"}],"recommendation":"major_revision","confidential_remarks":"The chief risk is that the central efficiency claims rest on incomparable baseline measurements; I recommend that the editor require a re-measurement of all baselines under the same input resolution and hardware before accepting the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the architecture is a reasonable combination of known ideas, but the headline efficiency numbers don't survive a close look at the measurement protocol. The LeMoRe* row in Table 1 is the tell: it is marked as trained at 448x448 with 0.6 GFLOPs, while the footnote says all GFLOPs are for 512x512 input. 0.8*(448/512)^2 = 0.61, so the GFLOPs were actually computed at 448, not 512. Either the footnote or the number is wrong, and that casts doubt on every derived reduction percentage in Section 4.1.\n\nWhat is genuinely new: the Cartesian encoder's three permuted 1x1 convolutions, the nested attention that computes all-pairs QK interactions across three QKV sets, and the gated fusion. Each piece is familiar in isolation—triplet attention, coordinate attention, multi-head attention—but I'm not aware of another lightweight segmentation paper that combines them this way. The ablation in Table 2 is informative: each extra view adds a bit, and nested attention gives a large jump, so the design choices are at least empirically motivated. Results on four datasets suggest the model is trainable and not a single-dataset fluke.\n\nSoft spots, in order of severity. First, the comparability problem above. Tables 3-5 have no resolution note, and the baseline GFLOPs look like they were pulled from papers at different input sizes (PSPNet at 423 GFLOPs is not at 1024x512). Second, no error bars or multiple seeds; every number is a single run. Third, Table 6 reports object detection results with no methodology—which RetinaNet settings, which dataset, what input resolution? Fourth, Section 3.1 asserts that the Cartesian views are orthogonal and maximize joint entropy, with no proof or citation; it is not obvious that 1x1 convolutions on permuted feature maps produce orthogonal subspaces. Fifth, the related work misses the closest attention variants, which weakens the novelty framing but does not kill it.\n\nThe stress-test note is right, and the central trade-off claim is not yet supported. Still, the architecture idea deserves a serious referee. The paper should go to review, with the expectation of major revisions: re-benchmarking under one protocol, adding error bars, documenting the detection experiment, and either proving or dropping the orthogonality/entropy claims. As it stands, I would not cite the efficiency numbers, but I'd flag the paper for anyone working on lightweight segmentation.","headline":"A plausible architecture, but the headline efficiency trade-off rests on ill-matched baselines and a directly contradicting footnote.","tokens_in":11141,"tokens_out":3531,"would_cite":false,"duration_ms":29903,"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 LeMoRe reaches 33.5% mIoU on ADE20K with 0.8 GFLOPs and 1.6M parameters, balancing accuracy and compute via explicit and implicit multiview modeling.","keywords":["lightweight semantic segmentation","Cartesian encoder","nested attention","multiview feature modeling","efficient vision transformer","ADE20K","CityScapes","gated fusion"],"falsifier":"Run LeMoRe and, say, LR-ASPP on the same device at 512x512 and measure end-to-end latency and peak memory; if LeMoRe's latency is not substantially lower than the baseline, the efficiency claim is undercut. Also, ablate the nested attention by replacing it with a single global attention of the same parameter budget; if ADE20K mIoU does not drop by roughly 5 points, the claimed mechanism is not the source of the gain.","tokens_in":9969,"feed_emoji":"⚡","tokens_out":6204,"duration_ms":55486,"temperature":0.7,"pith_summary":"The paper sets out to show that a lightweight semantic segmentation network can match the accuracy of much heavier transformer-based models by replacing a monolithic encoder with a decomposition-modeling-reconstruction scheme. Features are split into explicit Cartesian views, aligned to the height, width, and depth axes, plus implicit views learned through nested attention over query-key pairs. On the ADE20K benchmark the model reports 33.5% mIoU at 0.8 GFLOPs, 1.6M parameters, and 24 ms latency, which the authors claim outperforms comparable lightweight solutions in the efficiency-accuracy trade-off. The same design is shown to transfer to CityScapes, Pascal Context, COCO-Stuff, and object detection.","feed_headline":"Lightweight segmenter hits 33.5% mIoU at 0.8 GFLOPs","feed_subtitle":"Explicit Cartesian views plus nested attention claim transformer-level accuracy at a fraction of compute.","key_machinery":"The load-bearing mechanism is the Cartesian Encoder: three 1x1 convolutions applied to feature-map permutations along height, width, and depth, yielding transverse, frontal, and lateral views that are summed and refined with depthwise convolutions and channel attention. Complementing it is the Nested Attention module, which forms three query-key-value sets, computes attention maps between every query and every key across the sets, aggregates these into one map, and applies it to all value sets before a feed-forward network. Together they produce the explicit and implicit views that the Gated Fusion Module combines.","core_discovery":"The central claim is that modeling a feature map through multiple complementary views—three predefined Cartesian directions and three learned implicit views—captures local and global context with lower cost than a full transformer encoder. The Cartesian Encoder permutes the feature map along each spatial axis and applies 1x1 convolutions to produce transverse, frontal, and lateral views; the Nested Attention module computes all pairwise query-key attention maps across three sets and aggregates them to weight three value sets. A gated fusion module then recombines the local explicit and global implicit features. The paper reports that each added view contributes a measured mIoU gain (from 27.2% with one view to 33.5% with all three views, channel attention, and nested attention) while GFLOPs stay under 0.81 and parameters at 1.60M on ADE20K. It further reports consistent efficiency advantages on three other segmentation datasets and a detection transfer.","pith_inferences":["The ablation pattern (nested attention alone gives +5.0 mIoU, the Cartesian views add about +1.5) suggests the accuracy gain is driven more by the implicit global-attention term than by the explicit Cartesian decomposition.","The same decomposition-modeling-reconstruction pattern could be applied to other dense prediction tasks such as depth estimation, where explicit directional structure may also help.","A direct same-hardware latency comparison against LR-ASPP and U-MixFormer would settle whether the 24 ms figure reflects the architecture or favorable implementation; the paper does not provide re-measured latencies for baselines.","Because the nested attention computes all cross-products of three query-key sets, its cost grows quadratically with the number of implicit views; the choice of three is a design assumption that may not be optimal for other resolutions."],"forward_implications":["If the reported numbers hold, lightweight segmentation can reach roughly 33% mIoU on ADE20K at under 1 GFLOP, a point on the accuracy-efficiency curve that transformer encoders do not currently occupy.","The ablation results imply that each additional Cartesian view costs only about 0.02 to 0.04 GFLOPs while adding 0.4 to 1.5 mIoU, making multiview decomposition a cheap source of representational gain.","The CityScapes result of 65.0% mIoU at 1.7 GFLOPs suggests the architecture is usable for real-time urban scene parsing, and the ADE20K result supports its use in resource-constrained settings.","The detection transfer experiment indicates the learned views generalize as a backbone beyond segmentation, improving mAP over ShuffleNetV2 and MobileNetV3 backbones at similar FLOPs."],"supporting_citations":[{"why":"The best-performing lightweight transformer baseline; the paper's headline efficiency gains (73.8% fewer parameters, 86.9% fewer GFLOPs) are computed relative to it.","marker":"[3]"},{"why":"The SegFormer MiT-B0 baseline used for quantitative comparison and for the visual segmentation examples.","marker":"[26]"},{"why":"The LR-ASPP lightweight baseline that LeMoRe improves upon by +1.0 mIoU while cutting GFLOPs by 38.5%.","marker":"[30]"},{"why":"The ADE20K dataset on which the main accuracy numbers and all ablations are reported.","marker":"[4]"},{"why":"The CityScapes dataset used to support the real-time urban segmentation claim.","marker":"[5]"},{"why":"ImageNet pre-training, which the paper states is applied to all models for fair comparison.","marker":"[35]"}],"fun_headline_variants":["LeMoRe: 33.5% mIoU at 0.8 GFLOPs with nested attention","Cartesian views plus implicit modeling cut compute to 0.8 GFLOPs","Nested attention lets lightweight segmenter hit 33.5% mIoU","Multi-view attention matches transformer accuracy at 0.8 GFLOPs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central trade-off claim assumes that GFLOPs, latency, and parameter counts for all compared baselines are measured under the same protocol, meaning the same input resolution, hardware, and inference framework; in the paper, latency is reported only for LeMoRe and most baseline costs come from their original publications.","fun_headline_variants_meta":{"raw":{"variants":["LeMoRe: 33.5% mIoU at 0.8 GFLOPs with nested attention","Cartesian views plus implicit modeling cut compute to 0.8 GFLOPs","Nested attention lets lightweight segmenter hit 33.5% mIoU","Multi-view attention matches transformer accuracy at 0.8 GFLOPs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000832,"raw_usage":{"total_tokens":3590,"prompt_tokens":863,"completion_tokens":2727,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":479,"completion_tokens_details":{"reasoning_tokens":2632}},"tokens_in":479,"tokens_out":2727,"duration_ms":19203,"temperature":1.0,"reasoning_tokens":2632,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:54:08.023908+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run LeMoRe and, say, LR-ASPP on the same device at 512x512 and measure end-to-end latency and peak memory; if LeMoRe's latency is not substantially lower than the baseline, the efficiency claim is undercut. Also, ablate the nested attention by replacing it with a single global attention of the same parameter budget; if ADE20K mIoU does not drop by roughly 5 points, the claimed mechanism is not the source of the gain.","supporting_citations":[{"cited_title":"Segformer: Simple and efficient design for semantic segmentation with transformers,","cited_arxiv_id":null,"evidence_quote":"The SegFormer MiT-B0 baseline used for quantitative comparison and for the visual segmentation examples."},{"cited_title":"Searching for mobilenetv3,","cited_arxiv_id":null,"evidence_quote":"The LR-ASPP lightweight baseline that LeMoRe improves upon by +1.0 mIoU while cutting GFLOPs by 38.5%."},{"cited_title":"The cityscapes dataset for semantic urban scene understanding,","cited_arxiv_id":null,"evidence_quote":"The CityScapes dataset used to support the real-time urban segmentation claim."},{"cited_title":"Imagenet: A large-scale hierarchical im- age database,","cited_arxiv_id":null,"evidence_quote":"ImageNet pre-training, which the paper states is applied to all models for fair comparison."}],"review_version":1}