{"id":"4f0a5a24-71ea-4a01-9c92-e0a9ef6d1a0a","arxiv_id":"1908.09443","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Sharing one 3x3 kernel across parallel atrous-convolution branches improves semantic segmentation accuracy while cutting parameters, relative to ASPP.","lead":"This paper proposes a convolutional module called KSAC in which one small kernel is reused across several parallel branches with different dilation rates, so the same weights see the input at multiple scales. The authors report higher segmentation accuracy on two benchmarks with fewer parameters than the standard ASPP module used in DeepLab networks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported mIOU gain is not isolated from the parameter reduction: KSAC changes both kernel-sharing and model capacity at once, so the central claim that sharing itself improves accuracy is not yet supported.","rationale":"I agree with the reader's weakest-assumption analysis: the central empirical claim rests on a comparison that confounds the sharing mechanism with a substantial reduction in parameters. The paper does not report a capacity-matched ablation, does not provide code, and reports single-run numbers without error bars, so the unique contribution of kernel-sharing is not isolated. This is the single most load-bearing concern because if the gain is merely a regularization effect of fewer parameters, the paper's explanatory narrative about generalization and representation ability collapses, even though the architecture might still be a useful practical module. I considered whether the parameter reduction itself is enough to justify the claim as an efficiency-accuracy trade-off, but the paper explicitly claims that sharing 'boosts generalization and representation abilities,' which requires the mechanism to be active. I also reviewed the wider-context argument in Section 5.3: the gain from adding rates (1,24) without parameters is consistent with sharing but again lacks a same-setup ASPP baseline with the same inference strategy. The reader's CONDITIONAL verdict is appropriate: the idea is plausible and the reported numbers are promising, but the missing control means the central claim should be conditionally accepted pending an ablation that isolates sharing from capacity. My verdict recommendation is therefore UNCHANGED.","tokens_in":11334,"tokens_out":5205,"duration_ms":57517,"concrete_test":"Train an ASPP control on PASCAL VOC 2012 val under the same protocol as Table 1, with three separate 3x3 kernels but output channels per branch divided by 3 so the total number of 3x3 weights equals the single shared kernel in KSAC. Run at least three seeds for both this control and KSAC, and compare mean mIOU and variance. If the capacity-matched ASPP matches KSAC within seed noise, kernel-sharing is not the active ingredient; if KSAC still wins by a margin exceeding the seed spread, the sharing mechanism itself is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing claim is that sharing one 3x3 kernel across parallel atrous branches improves segmentation accuracy beyond the usual ASPP design. The evidence in Table 1, however, changes at least two variables simultaneously: (1) the kernel-sharing constraint itself and (2) the total parameter count (54.3M vs. 44.8M for Xception). Section 3.2 attributes the improvement to sharing, arguing that the shared kernel sees more effective training samples because small and large objects train different branches in ASPP. That argument is questionable: in both ASPP and KSAC, every training image is fed through every parallel branch, so every sample already contributes gradients to every branch kernel. The paper provides no ablation that matches capacity while keeping kernels separate, e.g., an ASPP variant with three separate 3x3 kernels but output channels reduced per branch so that total 3x3 parameters equal the single shared kernel. Without such a control, the observed mIOU gain could be driven by reduced capacity acting as implicit regularization, by different optimization dynamics, or by implementation differences, rather than by the kernel-sharing mechanism itself. The lack of released code or error bars further prevents ruling out run-to-run or implementation variation. Thus the central conceptual claim, that sharing kernels improves generalization and representation, is not yet securely established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Kernel-Sharing Atrous Convolution (KSAC), a modification of Atrous Spatial Pyramid Pooling (ASPP) in which a single 3x3 kernel is shared across parallel atrous branches with different rates, so that the same kernel processes the input feature maps multiple times with different receptive fields. The authors argue that this sharing improves the generalization and representation ability of the kernels, increases the effective number of training samples, reduces model parameters, and allows the network to benefit from a wider range of atrous rates. Experiments on PASCAL VOC 2012 and ADE20K report consistent mIOU improvements over the ASPP baseline with reduced parameter counts: for example, with an Xception backbone on VOC 2012 val, mIOU rises from 83.34% (DeepLabV3+/ASPP) to 85.96% (KSAC) while parameters drop from 54.3M to 44.8M, and extending the rate set to (1,6,12,18,24) further improves mIOU to 87.01%. The paper also reports 88.1% on the VOC 2012 test set and 45.47% on ADE20K validation.","tokens_in":11559,"tokens_out":3101,"duration_ms":30145,"significance":"If the reported gains are attributable to the kernel-sharing mechanism itself, KSAC is a simple, parameter-efficient drop-in replacement for ASPP that could be broadly useful in semantic segmentation and other dense prediction tasks. The empirical comparisons against public benchmarks are appropriate for the claim, and the reported improvements are consistent across two backbones and a second dataset, which strengthens the plausibility. However, the central empirical claim is not yet isolated from confounding factors: the main comparison changes both the sharing constraint and the total parameter count simultaneously, and the paper provides no ablation that matches capacity while keeping kernels separate, no error bars or multiple-seed results, and no released code. The conceptual explanation in Section 3.2 is asserted rather than measured, and the speed/memory comparison with prior work is not controlled for output stride. Thus the result is promising but the causal attribution to sharing is not yet securely established.","major_comments":[{"comment":"The primary comparison between KSAC and ASPP changes at least two variables at once: the weight-sharing constraint and the total number of parameters (54.3M vs 44.8M for Xception). Consequently, the reported mIOU gain could be caused by reduced capacity acting as regularization, by different optimization dynamics, or by implementation differences rather than by the sharing mechanism itself. Please add a capacity-matched control, such as an ASPP variant with separate per-branch 3x3 kernels but output channels per branch reduced so that the total 3x3 parameter budget equals the single shared kernel, and report mIOU and parameter counts for both variants.","section":"Table 1, Section 5.1"},{"comment":"The claim that sharing increases the number of effective training samples is not supported by the paper's own setup. In ASPP, every training image is already passed through every parallel branch, so every image contributes gradients to each branch kernel; the assertion that small or large objects are only effective for training the branch with a matching atrous rate is not substantiated. Please either remove this explanation or support it with direct evidence, for example gradient statistics per branch, or an ablation that isolates the sharing effect.","section":"Section 3.2"},{"comment":"The comparison with prior work is not controlled for output stride: the KSAC results in Table 3 are obtained with OS=16, while the competing results from DeepLab V3, DeepLab V3+, and EMA are obtained with OS=8. The paper also claims in Section 5.4 that KSAC achieves similar segmentation results under OS=8 and OS=16, but no OS=8 experiment is reported anywhere. This weakens both the accuracy comparison and the speed/memory claims, and the relevant experiment or a clear caveat should be added.","section":"Sections 5.3 and 5.4, Tables 2 and 3"},{"comment":"No error bars, multiple-seed results, or statistical significance tests are reported for any of the mIOU differences. Given that the MobileNetV2 improvement is only 0.6 percentage points (75.70% vs 76.30%), run-to-run and implementation variation could plausibly account for the gap. Please report variance across at least three runs for the key comparisons, or otherwise justify that the differences are stable.","section":"Section 5.1 and Table 1"}],"minor_comments":[{"comment":"There are typos and formatting issues, such as 'whe compared' in the abstract and 'Tensorflow2.0' in Section 4.2; these should be corrected.","section":"Abstract and Introduction"},{"comment":"The notation KERNEL(shape) is not defined; please specify how the shared kernel is initialized and whether it is updated by gradients from all branches jointly.","section":"Algorithm 1"},{"comment":"The sentence 'when the Xception decoder is used, about ten times of parameters have been reduced compared with MobileNetV2' is unclear; it likely refers to the Xception encoder, and the magnitude should be stated more precisely.","section":"Section 5.2"},{"comment":"The feature-map visualization is descriptive but not quantitative; consider adding a numerical measure, such as gradient statistics or feature discriminability, to support the claim of clearer edges and contours.","section":"Figure 5"},{"comment":"Several references are incomplete or inconsistently formatted, for example [14] lists 'International Conference on Computer Vision and Pattern Recognition' where the venue should be ICCV, and some arXiv identifiers appear without full citation details.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The idea is simple and potentially useful, and the reported numbers are consistently in its favor, but the paper currently does not distinguish the sharing mechanism from the parameter reduction. The requested capacity-matched ablation is essential and should be the focus of the revision. I also note that the abstract promises code release, but no code or detailed training configuration is provided; making it available would substantially increase confidence in the empirical claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe genuinely new thing here is the module: one 3×3 kernel shared among parallel atrous branches, so the same weights are applied at multiple dilation rates. That is a clean, simple idea I don't recall seeing before in the ASPP literature, and the parameter savings are a direct consequence of the design. The reported numbers also look encouraging: 85.96 vs 83.34 mIOU on VOC val with Xception, and 45.47 on ADE20K. If those hold under controlled comparison, KSAC is a useful drop-in replacement, especially for mobile.\n\nThe soft spot is exactly where the reader put it: the comparison changes two variables at once, sharing and capacity. Table 1 drops parameters from 54.3M to 44.8M (17%), and there is no ablation that matches capacity while keeping kernels separate—say, fewer output channels per branch—to show the gain comes from sharing rather than from the regularization effect of a smaller model. Section 3.2's \"effective training samples\" argument also does not survive scrutiny: in ASPP every image already passes through every branch, so every branch kernel already receives gradients from all objects. Sharing does not increase the number of training samples; it ties gradients across branches, which is more like weight tying. That could be a legitimate regularizer, but the paper does not test it as such. The wider-rate comparison (1,6,12,18,24) also lacks a same-setup ASPP baseline, so the claim that KSAC uniquely benefits from larger rates is asserted, not shown.\n\nMinor but compounding: no code despite the abstract promising release, no error bars or seeds, and the OS=16 vs OS=8 speed comparison in Section 5.4 compares against different training/eval setups. None of these are fatal alone, but together with the main confound they leave the central mechanism unsupported.\n\nBottom line: the idea deserves a serious referee. It is a plausible architectural variation with a clear efficiency advantage, and the accuracy claim may well be real. But the current manuscript does not isolate the mechanism. Peer review should demand a capacity-matched ablation, a same-setup rate=24 ASPP control, and code before acceptance.","headline":"Novel kernel-sharing atrous module with a clear efficiency gain, but the reported accuracy improvement is confounded with parameter reduction and the stated training-sample mechanism does not hold up.","tokens_in":12082,"tokens_out":3101,"would_cite":false,"duration_ms":33613,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that sharing one 3x3 kernel across parallel atrous branches at different rates lets a segmentation network see each feature map multiple times, improving accuracy while reducing parameters, and that this shared-kernel…","keywords":["semantic segmentation","atrous convolution","dilated convolution","kernel sharing","multi-scale context","ASPP","PASCAL VOC 2012","ADE20K"],"falsifier":"Train a comparison model in which each atrous branch keeps its own 3x3 kernel but with output channel count reduced so total parameters match KSAC; if this model matches or exceeds KSAC's mIOU, the gain is not caused by sharing the kernel.","tokens_in":11109,"feed_emoji":"🖼️","tokens_out":7117,"duration_ms":62606,"temperature":0.7,"pith_summary":"This paper tries to show that a single 3x3 kernel, shared across parallel atrous-convolution branches with different dilation rates, can outperform the standard ASPP module in semantic segmentation while using far fewer parameters. The idea is that the same kernel sees each feature map multiple times at different receptive fields, so the network gains multi-scale context and every object contributes to training one kernel rather than one branch-specific kernel. On PASCAL VOC 2012 with Xception, KSAC raises mIOU from 83.34% to 85.96% and saves about 10M parameters; with rates extended to (1,6,12,18,24), mIOU reaches 87.01%. If correct, KSAC offers a simple drop-in replacement for ASPP that is smaller, faster, and better able to exploit wider context.","feed_headline":"One kernel sees many receptive fields: accuracy hits 85.96%","feed_subtitle":"One 3x3 kernel shared by parallel atrous branches replaces ASPP, lifting accuracy on VOC and ADE20K.","key_machinery":"The central object is the kernel-sharing atrous convolution module: one 3x3 kernel reused by parallel atrous branches at different rates (typically 6, 12, 18, and optionally 1 and 24), with batch normalization and ReLU applied to each branch output before concatenation together with image-level features. This carries the argument because it makes the module's parameter count constant in the number of branches rather than linear, and because every object in the training images contributes to training the single shared kernel at every receptive field.","core_discovery":"The paper's central claim is that replacing the parallel atrous branches of ASPP, each with its own 3x3 kernel, with a single 3x3 kernel shared across the same branches improves semantic segmentation accuracy and reduces model size. The shared kernel sees the same feature map at multiple atrous rates, which the paper argues increases each kernel's effective training samples and lets it learn both local detail and global context. On PASCAL VOC 2012 validation with an Xception backbone, KSAC raises mIOU from 83.34% to 85.96% while cutting parameters from 54.3M to 44.8M; extending the rates to (1,6,12,18,24) raises mIOU further to 87.01% without adding parameters. On the VOC test set KSAC reaches 88.1% mIOU, and on ADE20K validation it reaches 45.47% mIOU under multi-scale evaluation.","pith_inferences":["A direct test of the paper's 'increased effective training samples' explanation would be to compare per-class accuracy between KSAC and ASPP: the sharing hypothesis predicts larger gains for classes with extreme object scales, not uniform gains.","The sharing mechanism is not specific to segmentation; a similar single-kernel multi-rate design could replace dilated parallel branches in object detection or instance segmentation heads, where multi-scale context also matters.","A capacity-matched ASPP variant with reduced per-branch channels would separate the contribution of sharing from the contribution of having fewer parameters, and would clarify whether the mechanism itself or the regularization effect of fewer parameters drives the reported gains."],"forward_implications":["With the Xception backbone on PASCAL VOC 2012 validation, replacing ASPP with KSAC raises mIOU from 83.34% to 85.96% and cuts model size from 54.3M to 44.8M parameters.","With MobileNetV2, KSAC reduces parameters by about one third (4.5M to 3.0M) while improving mIOU from 75.70% to 76.30%.","Adding rates 1 and 24 to KSAC improves mIOU from 85.96% to 87.01% with no additional parameters, whereas ASPP degrades when rate 24 is added.","KSAC reports 88.1% mIOU on the PASCAL VOC 2012 test set and 45.47% mIOU on ADE20K validation with multi-scale evaluation, under output stride 16.","Because the shared kernel is reused by every branch, adding more atrous branches does not increase the module's parameter count, so wider context can be explored at no memory cost in the convolution weights."],"supporting_citations":[{"why":"Supplies the DeepLabV3+ architecture that KSAC replaces as the baseline it must beat.","marker":"[5]"},{"why":"Provides the standard ASPP rate setting (6,12,18) and the observation that adding rate 24 slightly hurts ASPP.","marker":"[4]"},{"why":"Introduces the multi-branch atrous convolution idea in DeepLab that KSAC modifies.","marker":"[2]"},{"why":"Provides the PSPNet multi-pooling baseline and comparison on both VOC and ADE20K.","marker":"[30]"},{"why":"Supplies the SBD trainaug dataset used for VOC training and evaluation.","marker":"[8]"},{"why":"Provides the MobileNetV2 backbone used in the lightweight experiments.","marker":"[17]"},{"why":"Supplies the ADE20K dataset used for the additional segmentation benchmark.","marker":"[32]"}],"fun_headline_variants":["One shared kernel, many receptive fields: 85.96% mIOU","Kernel-sharing atrous convolution cuts params, boosts mIOU","Share one kernel across atrous branches for free accuracy","KSAC: single kernel, multiple views, 85.96% mIOU","One kernel sees all: atrous sharing lifts mIOU to 85.96%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the accuracy gains come from the sharing mechanism itself rather than from the accompanying reduction in parameters or other differences between the compared setups.","fun_headline_variants_meta":{"raw":{"variants":["One shared kernel, many receptive fields: 85.96% mIOU","Kernel-sharing atrous convolution cuts params, boosts mIOU","Share one kernel across atrous branches for free accuracy","KSAC: single kernel, multiple views, 85.96% mIOU","One kernel sees all: atrous sharing lifts mIOU to 85.96%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000288,"raw_usage":{"total_tokens":1741,"prompt_tokens":1048,"completion_tokens":693,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":664,"completion_tokens_details":{"reasoning_tokens":607}},"tokens_in":664,"tokens_out":693,"duration_ms":5662,"temperature":1.0,"reasoning_tokens":607,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:11:18.307184+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a comparison model in which each atrous branch keeps its own 3x3 kernel but with output channel count reduced so total parameters match KSAC; if this model matches or exceeds KSAC's mIOU, the gain is not caused by sharing the kernel.","supporting_citations":[{"cited_title":"Encoder-decoder with atrous separable convolution for semantic image segmentation","cited_arxiv_id":null,"evidence_quote":"Supplies the DeepLabV3+ architecture that KSAC replaces as the baseline it must beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the multi-branch atrous convolution idea in DeepLab that KSAC modifies."},{"cited_title":"Pyramid scene parsing network","cited_arxiv_id":null,"evidence_quote":"Provides the PSPNet multi-pooling baseline and comparison on both VOC and ADE20K."},{"cited_title":"Semantic contours from inverse detectors","cited_arxiv_id":null,"evidence_quote":"Supplies the SBD trainaug dataset used for VOC training and evaluation."},{"cited_title":"Mobilenetv2: Inverted residuals and linear bottlenecks","cited_arxiv_id":null,"evidence_quote":"Provides the MobileNetV2 backbone used in the lightweight experiments."},{"cited_title":"Scene parsing through ade20k dataset","cited_arxiv_id":null,"evidence_quote":"Supplies the ADE20K dataset used for the additional segmentation benchmark."}],"review_version":1}