{"id":"5c154d6a-0b47-49b4-842a-89814318ed0e","arxiv_id":"2411.09838","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A U-Net variant using channel-wise 1D convolutions and pixel-shuffle operations substantially reduces model size, but the claim of accuracy preservation is contradicted by the paper's own results on general datasets.","lead":"OneNet replaces 2D convolutions in a U-Net with pixel-unshuffle downscaling and channel-wise 1D convolutions, cutting parameters by up to 71 percent. The paper says accuracy is preserved, but its own tables show major accuracy drops on general segmentation datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own Table 1 contradicts the central accuracy-preservation claim: OneNete,4 drops mIOU by 12–32% on VOC and Oxford Pet, not \"within 1%\" as claimed in Sec 4.1.","rationale":"I agree with the reader's REJECT verdict, but the single most load-bearing concern is not the pixel-unshuffle premise that the reader lists as the weakest assumption. Even if pixel-unshuffle does transfer spatial information to the channel axis, the paper's own results contradict the advertised claim that accuracy is preserved: Table 1 shows large, consistent drops on PASCAL VOC and Oxford Pet, while Sec 4.1 states the drop is within 1%. This is an internal inconsistency, not a disagreement with an external consensus, and it directly undermines the abstract's \"preserves accuracy effectively\" claim. The FLOPs mismatch in Table 2 (256x256 for baselines vs 512x512 for OneNet) is a second independent problem that weakens the efficiency argument, but it is not needed to reject the central accuracy claim. The reader's rationale does mention the accuracy contradiction, so my agreement is partial rather than full, but their formal weakest assumption is a different (and less decisive) issue. A controlled multi-seed reproduction on the two hardest general datasets would settle whether the observed drops are robust or an artifact of a single run; with the current single-seed numbers, the central claim is unsupported.","tokens_in":11581,"tokens_out":7560,"duration_ms":76510,"concrete_test":"Reproduce the PASCAL VOC and Oxford Pet rows of Table 1 using the released OneNet code, training OneNete,4, OneNeted,4, and U-Net4 with identical hyperparameters and 5 random seeds, and report mean plus 95% CI mIOU; if the mean relative mIOU drop exceeds 1%, the Sec 4.1 claim is falsified by a controlled rerun.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that OneNet \"preserves accuracy effectively\" while reducing parameters (abstract), and Sec 4.1 sharpens this to \"maintaining an accuracy drop within 1% across datasets.\" This is falsified by the paper's own Table 1 on general segmentation data. Against the U-Net4 baseline, OneNete,4 mIOU drops from 0.182 to 0.160 on PASCAL VOC (~12% relative), from 0.316 to 0.216 on Oxford Pet full-mask (~32% relative), and from 0.713 to 0.636 on the 3-class Pet version (~11% relative). The full encoder-decoder OneNeted,4 does worse (e.g., 0.149 vs 0.182 on VOC and 0.535 vs 0.713 on Pet-small). Absolute mIOU differences are also 2–10 points, so no reading of \"within 1%\" is consistent with the reported numbers. The paper acknowledges an 11–15% drop in Sec 4.1, which directly contradicts the unqualified abstract claim. The FLOPs comparison in Table 2 is also compromised because baselines use 256x256 inputs while OneNet uses 512x512, so the efficiency claim cannot rescue the accuracy loss. The pixel-unshuffle premise may be plausible, but the empirical accuracy-preservation claim fails on the submitted evidence before that premise needs to be adjudicated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes OneNet, a U-Net variant that replaces the standard 2D convolutions in the encoder (or encoder-decoder) with channel-wise 1D convolutions combined with pixel-unshuffle and pixel-shuffle operations. The central claim is that this architecture preserves segmentation accuracy while reducing parameters by up to 47% (encoder-only) or 71% (fully 1D encoder-decoder). The method is evaluated on PASCAL VOC, Oxford Pet with two mask versions, and three Medical Segmentation Decathlon datasets (Heart, Brain, Lung), reporting cross-entropy loss, mAP0.5, mIOU, and Dice scores. The paper also reports parameter counts, FLOPs, and memory usage, and includes an ablation on adding spatial convolutions.","tokens_in":11852,"tokens_out":4354,"duration_ms":43313,"significance":"If the central claim were true, OneNet would be a practically useful edge-deployable segmentation backbone, and the combination of pixel-unshuffle with channel-wise 1D convolutions would be an interesting architectural contribution. The paper has some strengths: the code is released, the evaluation covers six datasets, and the ablation on spatial convolutions is a relevant control. However, the main empirical claim is not supported by the paper's own results: Table 1 shows accuracy drops far larger than the 'within 1%' claimed in Section 4.1. In addition, the theoretical derivation in Section 3.1 is built on a garbled equation and an inconsistent algorithm, and the FLOPs comparison in Table 2 uses different input resolutions for the baselines and for OneNet. The architectural premise that pixel-unshuffle moves spatial information into the channel dimension is plausible but not formally justified, and the paper's own measurements do not demonstrate accuracy preservation on general segmentation tasks. The significance of the contribution, as submitted, is therefore not established.","major_comments":[{"comment":"The claim 'OneNet achieves model size reduction while maintaining an accuracy drop within 1% across datasets' is directly contradicted by the paper's own numbers. For OneNete,4 against the U-Net4 baseline, mIOU drops from 0.182 to 0.160 on PASCAL VOC (about 12% relative), from 0.316 to 0.216 on Oxford Pet full-mask (about 32% relative), and from 0.713 to 0.636 on the 3-class Pet version (about 11% relative). The encoder-decoder variant OneNeted,4 performs worse still, e.g., 0.149 vs 0.182 on VOC and 0.535 vs 0.713 on Pet-small. The same section later acknowledges 'an 11% and 15% drop in accuracy' and 'a 10% decrease', which is inconsistent with the unqualified 'within 1%' statement. Since the abstract's central claim is that OneNet 'preserves accuracy effectively', this is a load-bearing empirical contradiction that cannot be fixed by rewriting a single sentence.","section":"Section 4.1, Table 1"},{"comment":"The FLOPs and efficiency comparison is invalid because the baselines and OneNet are evaluated at different input resolutions. The text states that a sample tensor of size (1,3,256,256) is used for baselines and (1,3,512,512) for OneNet, while the Table 2 caption says a tensor of size (1,3,256,256) was used as the network input. With OneNet receiving a 512x512 input and the baselines a 256x256 input, the reported FLOPs values are not comparable, and the claimed FLOPs reductions (25% for the encoder, 78% for the full encoder-decoder) are not supported. The parameter-count reduction is a separate quantity, but the computational-efficiency claim in the abstract and Section 4.2 relies on this flawed comparison.","section":"Section 4.2, Table 2"},{"comment":"The formal definition and the implementation of the proposed pixel-unshuffle downscaling are not internally consistent. Equation (1), D(X)_{i,a,b} = X_{⌊i/s2⌋,s a+⌊sa/i⌋,s b+sb(mod i)}, uses undefined quantities: s2 appears as a divisor but s is already defined as scale, ⌊sa/i⌋ divides by the channel index i, and the term sb(mod i) is not meaningful as written. Algorithm 1 declares an input of shape (B, C, H×W) and an output of shape (B, H×W×C), but the indexing loop and the reshape/transpose/flatten sequence do not produce that output: transpose(0,2,1,3).flatten(dim=1) on a (B,C, HW/4, 4) tensor yields shape (B, HW/4, 4C), not (B, H×W×C). The parameter-count derivation in Section 3.1 depends on this operation, so the theoretical justification for the efficiency gains is not presently verifiable.","section":"Section 3.1, Eq. (1), Algorithm 1"},{"comment":"The load-bearing architectural premise is that pixel-unshuffle transfers spatial knowledge to the channel axis, so that channel-wise 1D convolutions on the flattened tensor can replace 2D convolutions. This is asserted in Section 3.2 and again in Section 3.4 ('we can make a mild assumption that the channel dimension already consists of sufficient spatial information'), but it is not demonstrated formally or empirically. The ablation in Section 4.3 only shows that, under this specific design, adding a spatial convolution with kernel size 9 changes accuracy by at most about 2%; it does not show that the channel-wise 1D convolutions actually capture spatial relationships. The accuracy drops on VOC and Oxford Pet are consistent with the alternative hypothesis that the channel-wise 1D encoder loses spatial detail. A correct treatment would need either a formal argument about the information content of the unshuffled tensor or a comparison with a 2D-convolution control that uses the same number of channels, the same downsampling strategy, and the same training protocol.","section":"Section 3.2, Section 3.4, Section 4.3"}],"minor_comments":[{"comment":"The two paragraphs in Section 4.3 are near-duplicates; the second paragraph repeats the description of the ablation and the same reported finding, which makes the section read as if it was accidentally pasted twice.","section":"Section 4.3"},{"comment":"The shorthand PETF and PETS should be introduced in the caption more explicitly, since the text later refers to 'full-size Oxford Pet' and 'Oxford Pet dataset with fewer masks', which is easy to confuse with the two mask versions.","section":"Table 1"},{"comment":"Equation (1) should be rewritten with proper index notation and with clearly defined variables; as printed, the subscript ⌊sa/i⌋ and the term sb(mod i) cannot be parsed as a tensor-indexing operation.","section":"Section 3.1, Eq. (1)"},{"comment":"The sentence 'Receptiveness = S^L' uses a garbled superscript and would be clearer as 'receptive field scales as s^L, where s is the pixel-unshuffle scale and L is the number of layers'.","section":"Section 3.2"},{"comment":"The paper reports LCE (cross-entropy loss) in Table 1 but never analyzes it; either the column should be removed or its relevance to the accuracy claims should be discussed.","section":"Section 4.1"}],"recommendation":"reject","confidential_remarks":"The paper's most important claim, 'preserves accuracy effectively' and 'accuracy drop within 1%', is false on the evidence the paper itself reports. The theoretical section also contains an unusable equation and algorithm, and the FLOPs comparison is invalid. These are not local issues; they affect the paper's core contribution. I could see a revised version that reframes the contribution as an efficiency-accuracy trade-off rather than accuracy preservation, but that would require new experiments and a different central claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the architecture is a genuinely new combination and the parameter savings are real, but the headline claim doesn't survive contact with the paper's own Table 1. The accuracy drop on general segmentation is 11–32% relative, not \"within 1%\" as the abstract and Section 4.1 claim. The math in Section 3 is too garbled to verify.\n\nWhat's actually new: combining pixel-unshuffle downscaling with channel-wise 1D convolutions in a U-Net is something I haven't seen before. The parameter reductions (47% for the encoder, 71% for the full encoder–decoder) are plausible, and the code is provided. The ablation on spatial convolution is a reasonable sanity check, and the related work anchors to PixelShuffle and MobileNet properly.\n\nSoft spots: Section 4.1 says \"maintaining an accuracy drop within 1% across datasets\" but then concedes an 11% and 15% drop for VOC and Pet a few sentences later. Table 1 shows OneNete,4 dropping mIOU from 0.182 to 0.160 on VOC (12% relative), from 0.316 to 0.216 on Oxford Pet full-mask (32%), and from 0.713 to 0.636 on the small-mask version (11%). That's not within 1% by any reading. The FLOPs comparison in Table 2 uses 256×256 inputs for baselines and 512×512 for OneNet, so the efficiency claim is inflated. The theoretical derivation (Eq. 1, Algorithm 1) has garbled equations and inconsistent tensor shapes; I could not follow the arithmetic. The pixel-unshuffle premise—that spatial information is preserved in the channel axis—is asserted rather than analyzed. The ablation only shows that adding a spatial kernel doesn't help under this specific design, which is weak evidence for sufficiency.\n\nWhat's good: on the medical datasets the results are on par, so there may be a niche for this architecture in low-resolution or local-feature tasks. The paper does eventually acknowledge the trade-off, just not in the abstract or the bold claim.\n\nWho this is for: a researcher prototyping lightweight segmentation backbones might pick up the idea, but they'd need to re-implement and re-evaluate. Not for anyone relying on the abstract.\n\nRecommendation: I'd send it to peer review only if the authors commit to fixing the math, matching input resolutions, and rewriting the accuracy claims to match their own data. As submitted, the central claim is unsupported.","headline":"Genuinely new architecture combo with real parameter savings, but the accuracy-preservation claim is contradicted by the paper's own Table 1 and the math is too garbled to verify.","tokens_in":12431,"tokens_out":4378,"would_cite":false,"duration_ms":41104,"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":"OneNet swaps 2D convolutions for 1D channel-wise convolutions, shrinking U-Net by 47% while keeping segmentation accuracy.","keywords":["semantic segmentation","U-Net","1D convolution","pixel-unshuffle","pixel shuffle","model compression","edge deployment","channel-wise convolution"],"falsifier":"Take the OneNet encoder and replace the pixel-unshuffle permutation with a fixed random shuffle of the same channel geometry, keeping all other settings unchanged; if accuracy on a medical segmentation dataset does not drop, then the specific spatial organization claimed to carry spatial information is not what is doing the work. Alternatively, if a max-pooling baseline with the same channel counts matches OneNet's accuracy, the unshuffle's information-preservation benefit is not demonstrated.","tokens_in":11350,"feed_emoji":"🧩","tokens_out":5848,"duration_ms":50309,"temperature":0.7,"pith_summary":"This paper proposes OneNet, a U-Net variant that replaces 2D convolutions with channel-wise 1D convolutions, using pixel-unshuffle downscaling in the encoder and pixel-shuffle upscaling in the decoder to move spatial information into the channel axis. The central claim is that this design matches the segmentation accuracy of a standard U-Net while cutting parameters by up to 47% with only the encoder swapped, and by up to 71% in model size when the decoder is also replaced. The authors benchmark against U-Net, ResNet, and MobileNet backbones on six mask-prediction datasets, reporting accuracy drops within about 1% on medical tumor segmentation and larger drops on multi-class datasets such as PASCAL VOC and Oxford Pet. If correct, OneNet offers a drop-in efficient backbone for edge deployment, and the pixel-shuffle mechanism becomes a general template for replacing 2D spatial convolutions with cheaper 1D channel operations.","feed_headline":"OneNet swaps 2D for 1D convolutions, shrinking U-Net 47%","feed_subtitle":"Pixel-unshuffle moves spatial info into channels, preserving medical-mask accuracy within 1%.","key_machinery":"The load-bearing mechanism is the pixel-unshuffle downscaling operation, which reorganizes a tensor so that each spatial neighborhood of size $s \\times s$ becomes a set of extra channels, followed by channel-wise 1D convolutions of kernel size and stride $C_{\\text{in}}$ over the flattened $(B, HW C)$ tensor. Because the unshuffle moves spatial relationships into the channel axis, a 1D convolution along that axis can, in the authors' argument, capture the same information a 2D convolution would, at a fraction of the parameters. The decoder mirrors this with pixel-shuffle upscaling that moves information back from channels to space, and an optional spatial 1D convolution with a small kernel is included but shown by ablation to be nearly irrelevant. The paper also implements a 1D-compatible pixel-unshuffle algorithm so that all tensors can be kept in a single spatial dimension.","core_discovery":"OneNet's central discovery is that a semantic segmentation encoder does not need 2D convolutions if pixel-unshuffle first repackages spatial neighborhoods into the channel dimension. After this repackaging, a channel-wise 1D convolution over the flattened tensor can mix spatial and channel information with a parameter count proportional to $C_{\\text{in}} C_{\\text{out}}$ rather than $k^2 C_{\\text{in}} C_{\\text{out}}$, and the receptive field is set by the unshuffle scale rather than by kernel size. In the paper's own formulation, replacing a 2D block with a pixel-unshuffle plus 1D block reduces the per-block multiplication ratio by a factor of $k^2/2$, which translates to 47% fewer parameters for the encoder and a 71% smaller model for the fully 1D version. Across the six datasets, the paper reports that the encoder-only model stays within about 1% of U-Net's accuracy on medical tumor segmentation, while the fully 1D encoder-decoder trades a larger accuracy drop for the largest size reduction.","pith_inferences":["A natural extension is to test OneNet as a backbone inside diffusion models, where U-Net is often the computational bottleneck; the authors position the work as adaptable but do not run that experiment.","The larger accuracy drops on PASCAL VOC and Oxford Pet suggest the channel-wise 1D design may be better suited to local, low-class-count masks than to image-wide classification; a targeted experiment with high-resolution scenes containing large objects would clarify this.","The 1D-compatible pixel-unshuffle algorithm opens the door to using larger unshuffle scales or non-2x2 rearrangements to trade channel capacity for spatial context, which the paper does not explore.","One could test whether pretraining the OneNet encoder on ImageNet closes the gap to pretrained ResNet backbones; the paper deliberately does not pretrain, so this remains open."],"forward_implications":["OneNet's encoder can replace a standard U-Net encoder in existing pipelines, cutting parameters by roughly 47% while keeping medical-segmentation accuracy within about 1%.","A fully 1D encoder-decoder OneNet reaches a 71% smaller model, at the cost of larger accuracy losses on multi-class segmentation datasets.","Because the receptive field is controlled by the unshuffle scale rather than kernel size, model designers can tune for local feature detection without changing kernel dimensions.","The same pixel-unshuffle-plus-1D block pattern could be applied to other convolutional architectures beyond segmentation, including super-resolution and generative backbones.","The ablation suggests spatial convolutions contribute little in this design, so further size reductions are possible by removing them."],"supporting_citations":[{"why":"Provides the U-Net encoder-decoder architecture that OneNet modifies and the primary accuracy baseline.","marker":"[23]"},{"why":"Introduces PixelShuffle, the operation whose inverse (pixel-unshuffle) OneNet uses to move spatial information into the channel axis.","marker":"[24]"},{"why":"Supplies the depthwise-separable convolution comparison and the MobileNet block used in the spatial-convolution ablation.","marker":"[12]"},{"why":"Provides the ResNet backbones used as pretrained baselines in the segmentation comparisons.","marker":"[11]"},{"why":"Supplies the three Medical Segmentation Decathlon datasets (Heart, Brain, Lung) where OneNet reports accuracy within 1%.","marker":"[1]"},{"why":"Supplies the PASCAL VOC multi-class segmentation benchmark where the largest accuracy drop is observed.","marker":"[9]"},{"why":"Supplies the Oxford Pet breed-mask and subject-mask datasets used to test small-mask-count segmentation.","marker":"[20]"},{"why":"Gives prior evidence that networks built from directional 1D convolutions can match 2D ConvNets on ImageNet, motivating the 1D design.","marker":"[14]"}],"fun_headline_variants":["OneNet slims U-Net 47% with pure 1D convolutions","Drop 2D convs, keep accuracy: OneNet cuts U-Net by 47%","Pixel-unshuffle makes 2D convs obsolete: OneNet is 47% lighter","U-Net without 2D filters: OneNet shrinks params 47%","1D-only U-Net: OneNet matches accuracy, uses 47% fewer params"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole design rests on the idea that pixel-unshuffle moves spatial relationships into the channel axis, so a channel-wise 1D convolution can see them; if that repackaging does not actually preserve enough spatial information, the encoder loses detail the decoder cannot recover.","fun_headline_variants_meta":{"raw":{"variants":["OneNet slims U-Net 47% with pure 1D convolutions","Drop 2D convs, keep accuracy: OneNet cuts U-Net by 47%","Pixel-unshuffle makes 2D convs obsolete: OneNet is 47% lighter","U-Net without 2D filters: OneNet shrinks params 47%","1D-only U-Net: OneNet matches accuracy, uses 47% fewer params"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000942,"raw_usage":{"total_tokens":4031,"prompt_tokens":958,"completion_tokens":3073,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":2956}},"tokens_in":574,"tokens_out":3073,"duration_ms":21209,"temperature":1.0,"reasoning_tokens":2956,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:15:10.047494+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the OneNet encoder and replace the pixel-unshuffle permutation with a fixed random shuffle of the same channel geometry, keeping all other settings unchanged; if accuracy on a medical segmentation dataset does not drop, then the specific spatial organization claimed to carry spatial information is not what is doing the work. Alternatively, if a max-pooling baseline with the same channel counts matches OneNet's accuracy, the unshuffle's information-preservation benefit is not demonstrated.","supporting_citations":[{"cited_title":"Zhang, Shaoqing Ren, and Jian Sun","cited_arxiv_id":null,"evidence_quote":"Provides the ResNet backbones used as pretrained baselines in the segmentation comparisons."},{"cited_title":"The pascal visual object classes (voc) challenge","cited_arxiv_id":null,"evidence_quote":"Supplies the PASCAL VOC multi-class segmentation benchmark where the largest accuracy drop is observed."},{"cited_title":"Parkhi, Andrea Vedaldi, Andrew Zisserman, and C","cited_arxiv_id":null,"evidence_quote":"Supplies the Oxford Pet breed-mask and subject-mask datasets used to test small-mask-count segmentation."},{"cited_title":"Convolutional Networks with Oriented 1D Kernels","cited_arxiv_id":"2309.15812","evidence_quote":"Gives prior evidence that networks built from directional 1D convolutions can match 2D ConvNets on ImageNet, motivating the 1D design."}],"review_version":1}