{"id":"08b33db1-174c-4e4c-a95d-b6e365545fec","arxiv_id":"2501.15385","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DDUNet, a 0.33M-parameter U-Net with dynamic multi-scale convolution and dynamic classifier weights, reaches 95.3% accuracy for cloud segmentation on SWINySEG.","lead":"This paper presents DDUNet, a U-Net variant that adds dynamic multi-scale convolutions and a dynamic classifier head for cloud segmentation. It reports 95.3% accuracy on the SWINySEG dataset with only 0.33M parameters, targeting real-time use on small devices.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'superior performance in both accuracy and efficiency' claim is not supported: Table I shows MA-SegCloud beats DDUNet on every metric, and baseline split/training compatibility is unverified.","rationale":"The defensible kernel of the paper is empirical: a compact U-Net with dynamic modules reaches 95.3% accuracy on SWINySEG with 0.33M parameters. I reviewed the architecture and parameter math; nothing in Sec. III obviously makes the parameter count impossible. The soft spot is not the method itself but the evidential link between that number and the abstract's 'superior performance' claim. The paper's own Table I is enough to show DDUNet is not the most accurate method: MA-SegCloud is better on every row. Since the paper provides no error bars, the small remaining differences with CloudU-Net/DeeplabV3+ are uninterpretable as superiority. The lack of stated baseline training protocol and the absence of latency/FLOPS measurements make the efficiency claim unverifiable. A reproduction on the same split with seeded runs would settle all three points. This does not change the reader's conditional verdict; it sharpens the revision list.","tokens_in":8318,"tokens_out":9321,"duration_ms":83311,"concrete_test":"On the SWINySEG dataset, implement the exact 9:1 split described in Section IV.A, then retrain DDUNet, CloudU-Net, and MA-SegCloud with the same optimizer, epochs, batch size, and augmentations for at least three seeds. Report mean and std of accuracy/MIoU on the same test split, plus per-image inference time and FLOPs on identical hardware. This will settle (1) whether the baseline numbers are reproducible under the same protocol, (2) whether MA-SegCloud really beats DDUNet in accuracy, and (3) whether DDUNet's dynamic modules actually translate to lower latency than its parameter count implies.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table I internally contradicts the abstract. On every split, MA-SegCloud [38] reports higher accuracy, precision, mFβ, and MIoU than DDUNet; for day+night the gap is 96.9% vs 95.3% accuracy and 0.940 vs 0.884 MIoU. Therefore 'superior performance ... in both accuracy and efficiency' is false if accuracy is included. The only objective advantage shown is parameter count (0.33M vs 16.3M). Parameter count is not a direct measure of efficiency: DWBG generates per-sample weights in the decoder, and the paper reports no FPS, latency, or FLOPs, so the 'highly-efficient' title claim is unmeasured. Additionally, Section IV.A specifies DDUNet's split and training, but never states whether U-Net, PSPNet, DeeplabV3+, CloudSegNet, SegCloud, CloudU-Net, CloudU-Netv2, or MA-SegCloud were retrained under identical protocols or whether their Table I numbers were copied from the original papers. If the baselines used different splits, the accuracy-efficiency comparison is not a controlled experiment.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DDUNet, a U-Net variant for cloud segmentation that introduces two modules: Dynamic Multi-scale Conv2D (DMSC), which aggregates dilated depth-wise convolutions with input-dependent weights, and a Dynamic Weights and Bias Generator (DWBG), which generates per-sample classifier weights from pooled encoder and decoder features. The model is evaluated on the SWINySEG dataset under day-time, night-time, and day+night splits, achieving 95.3% accuracy and 0.884 MIoU on the day+night split with 0.33M parameters. The central claim is that DDUNet is both highly accurate and highly efficient, with a favorable accuracy-parameter tradeoff relative to existing cloud segmentation models.","tokens_in":8553,"tokens_out":2286,"duration_ms":20005,"significance":"If the reported results are reproducible and the comparison is fair, DDUNet would be a useful contribution to lightweight cloud segmentation: it is an order of magnitude smaller than DeeplabV3+, CloudU-Net, and MA-SegCloud while maintaining competitive accuracy. The code is publicly available, which strengthens reproducibility. However, the significance is currently undercut by overbroad claims in the abstract, unverified baseline comparability, and the absence of direct efficiency measurements such as latency or FLOPs.","major_comments":[{"comment":"The abstract claims that DDUNet 'achieve[s] superior performance over three different configurations of the SWINySEG dataset in both accuracy and efficiency.' This is directly contradicted by Table I, where MA-SegCloud [38] reports higher accuracy, precision, mFβ, and MIoU than DDUNet on every split (e.g., day+night accuracy 96.9% vs. 95.3%, MIoU 0.940 vs. 0.884). The only unambiguous advantage in the table is parameter count (0.33M vs. 16.3M). The claim should be revised to state that DDUNet offers comparable accuracy with far fewer parameters, rather than superior accuracy.","section":"Abstract and Table I"},{"comment":"The paper states that the authors follow [38] to split the SWINySEG dataset at 9:1, but it does not state whether the competing methods (U-Net, PSPNet, DeeplabV3+, CloudSegNet, SegCloud, CloudU-Net, CloudU-Netv2, MA-SegCloud) were retrained under this exact split and with the same training protocol, or whether the reported numbers were copied from their original papers. If the baselines used different splits or training settings, the accuracy and parameter-efficiency comparisons in Table I are not controlled, and the central efficiency-accuracy claim is not established. The authors need to clarify the baseline evaluation protocol and ideally retrain all baselines under identical conditions.","section":"Section IV.A (Experiments Setting)"},{"comment":"The title and text describe DDUNet as 'highly-efficient' and suitable for 'real-time' deployment, but the paper reports no FPS, inference latency, or FLOPs. Parameter count alone does not determine efficiency, and the per-sample weight generation in DWBG (Section III.C) may introduce runtime overhead that offsets the parameter savings. The paper should report direct efficiency measurements, preferably on the same hardware for all methods, before claiming real-time capability.","section":"Section IV.C (Quantitative Analysis)"},{"comment":"All quantitative results are based on a single train/test split with no error bars or multiple-seed runs. The differences between DDUNet and several baselines are small in some splits (e.g., day-time accuracy 95.3% vs. U-Net's 94.3%; night-time MIoU 0.900 vs. CloudU-Net's 0.912), and without variance estimates or statistical testing, it is not possible to determine whether these differences are meaningful. The paper should add results over multiple runs and report standard deviations or confidence intervals.","section":"Table I and Section IV.A"}],"minor_comments":[{"comment":"There is inconsistent naming between 'SWINSEG' (line in Introduction) and 'SWINySEG' (used elsewhere); please standardize to a single dataset name.","section":"Abstract and Introduction"},{"comment":"The terms 'reception field' should be 'receptive field', and 'Wights' (in 'Dynamic Wights and Bias Generator' and Figure 1c) should be 'Weights'.","section":"Throughout"},{"comment":"The paragraph beginning 'Recent research in remote sensing...' appears twice verbatim; one copy should be removed.","section":"Section II (Related Works)"},{"comment":"The ablation table is numbered 1, 2, 3, 5, 6, skipping 4. This appears to be a typo and should be renumbered consecutively.","section":"Table II"},{"comment":"The sentence 'We follow [38] to split the SWINySEG dataset ... with a ratio of 9:1 for training and testing' would benefit from specifying whether the split is stratified by day/night and how the day+night configuration is formed, since these choices affect comparability.","section":"Section IV.A"}],"recommendation":"major_revision","confidential_remarks":"The paper's core model is potentially useful and the parameter-accuracy tradeoff is interesting, but the current presentation overstates performance and under-specifies the experimental protocol. The issues are fixable within the manuscript's scope: temper the claims, document baseline retraining, add efficiency metrics, and report variance. I would be willing to review a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper's headline claim is wrong, but the underlying model is a legitimate lightweight segmentation net.\n\nThe new thing is a specific combination: a U-Net with dynamic multi-scale convolution (DMSC) and a dynamic weight/bias generator (DWBG). Both ideas are known in other forms — ASPP-style dilated branches with attention weighting, and hypernetwork-generated filters — but the combo for cloud segmentation with 0.33M params is not something I've seen. The ablation shows DMSC and DWBG each add real points over a 0.32M U-Net baseline, and the code is public. That's credit-worthy.\n\nThe problem is the abstract: 'superior performance ... in both accuracy and efficiency' is false on the paper's own numbers. In Table I, MA-SegCloud beats DDUNet on every metric on every split — day+night accuracy 96.9% vs 95.3%, MIoU 0.940 vs 0.884. So the accuracy claim is simply not supported. The efficiency claim is stronger but under-measured: the paper reports parameter count only, no FPS, latency, or FLOPs, and DWBG generates per-sample weights during inference, so 'highly-efficient' needs actual runtime numbers.\n\nThere's also a protocol question. Section IV.A says they follow [38] for the split, but never states whether the baselines were retrained under the same protocol or whether Table I numbers were taken from the original papers. If the baselines used different splits or training setups, the comparison isn't controlled. Also, all numbers are from a single train/test split with no error bars.\n\nThat said, the core result is plausible: a 0.33M model at 95.3% accuracy on SWINySEG is a useful engineering data point. The gap to MA-SegCloud is small while the parameter count is ~50x smaller. The paper just needs to (1) retract the accuracy-superiority wording, (2) add latency/FLOPs measurements, and (3) clarify or redo the baseline comparison.\n\nThis deserves a serious referee. The architecture and ablation are worth engaging with, and the overclaim is fixable. I'd send it to review with a request for major revision.","headline":"Abstract overclaims accuracy superiority, but the 0.33M-parameter model is a real lightweight segmentation contribution worth refereeing after fixes.","tokens_in":9093,"tokens_out":2107,"would_cite":false,"duration_ms":17497,"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":"A 0.33M-parameter U-Net variant reports 95.3% accuracy on cloud segmentation.","keywords":["cloud segmentation","U-Net","dynamic convolution","multi-scale feature extraction","depthwise convolution","lightweight network","SWINySEG","remote sensing"],"falsifier":"Run the released code on SWINySEG with the stated 9:1 split and the reported training schedule; the central claim fails if 95.3% accuracy and 0.884 MIoU do not reproduce, or if DeepLabV3+ and CloudU-Net, retrained under an identical protocol, match DDUNet's accuracy at a comparable parameter count.","tokens_in":8127,"feed_emoji":"☁️","tokens_out":4967,"duration_ms":42253,"temperature":0.7,"pith_summary":"The paper proposes DDUNet, a lightweight U-Net for separating cloud pixels from sky in ground-based camera images. It claims DDUNet reaches 95.3% accuracy and 0.884 mean IoU on the day+night SWINySEG configuration with only 0.33M parameters, placing it within about one point of much larger models such as DeepLabV3+ and CloudU-Net while using a fraction of their parameters. The central design idea is to replace fixed convolutions with two input-dependent modules: a dynamic multi-scale convolution that merges four dilated depthwise features with learned weights, and a dynamic weight-and-bias generator that creates per-image classification filters from encoder and decoder features. The paper positions this as a step toward accurate cloud segmentation that could run on resource-limited devices.","feed_headline":"0.33M-parameter U-Net hits 95.3% on cloud segmentation","feed_subtitle":"Dynamic multi-scale and per-image classifier weights keep accuracy near far larger models on SWINySEG.","key_machinery":"The two load-bearing modules are the Dynamic Multi-scale Conv2D (DMSC) and the Dynamic Weights and Bias Generator (DWBG). DMSC is a multi-branch block: after a 1x1 projection, four depthwise three-by-three convolutions with dilation rates 1 through 4 capture different receptive fields, a small linear-plus-softmax head computes per-image branch weights from global pooled features, and the weighted sum is fused by a short-cut connection. DWBG pools and concatenates encoder and decoder features, then runs two linear layers to generate the three-by-three weights and bias of the final classification convolution, so the classifier adapts to each input instead of using fixed kernels. Depthwise convolution throughout keeps the parameter count at 0.33M, and deep supervision on the last three decoder outputs with binary cross-entropy losses weighted 1, 0.5, and 0.2 drives convergence.","core_discovery":"The central claim is that binary cloud segmentation can be made dramatically cheaper without sacrificing accuracy by making the convolution kernels themselves conditional on the input. On SWINySEG, DDUNet reports 95.3% accuracy, 95.2% precision, 94.7% F-measure, and 0.884 MIoU on the combined day+night split, with 0.33M parameters; the ablation attributes most of the gain to the dynamic multi-scale convolution (93.0% to 94.8% accuracy at 0.28M parameters) and the remainder to the dynamic weight-and-bias generator (94.8% to 95.3% at 0.33M parameters). The authors claim superior accuracy-efficiency trade-offs across day-time, night-time, and day+night configurations, while noting that the raw accuracy of the largest baseline, MA-SegCloud, remains higher.","pith_inferences":["A natural extension the paper leaves untested is applying DDUNet to other all-sky datasets such as SWIMSEG and SWINSEG to see whether the input-dependent classifier generalizes across camera hardware and sky conditions.","Because DMSC computes branch weights from the input image itself, DDUNet should in principle adapt to scale distributions; this could be tested directly by comparing it against fixed-dilation U-Nets on images with unusually large or small cloud patches.","The efficiency ranking depends on the baseline numbers as reported; re-benchmarking all methods under a unified training protocol could shift the ordering, since the highest-accuracy baseline in Table I reports 96.9% accuracy with 16.3M parameters."],"forward_implications":["At 0.33M parameters and 95.3% accuracy on the day+night SWINySEG split, a model this small can plausibly run in low-latency settings on embedded devices for ground-based sky monitoring.","The ablation shows that swapping the baseline encoder for DMSC improves accuracy from 93.0% to 94.8% while reducing parameters to 0.28M, so multi-scale dynamic aggregation is the main source of the gain.","Adding DWBG raises accuracy from 94.8% to 95.3% at 0.33M parameters, demonstrating that per-input classifier weights add accuracy without adding many parameters.","Scaling the base channel count from 8 to 4 shrinks the model to 0.09M parameters with 93.6% accuracy, while scaling to 1.25M parameters yields only 95.4%, indicating that the chosen size sits near a useful efficiency frontier.","If the reported numbers reproduce, DDUNet offers a direct lightweight alternative to models like DeepLabV3+ and CloudU-Net that report similar accuracy with 1/9 to 1/100 of the parameter count."],"supporting_citations":[{"why":"Supplies the U-Net encoder-decoder architecture that DDUNet builds on.","marker":"[16]"},{"why":"Provides the CloudSegNet baseline with 0.005M parameters, the smallest model DDUNet compares against.","marker":"[5]"},{"why":"Provides the CloudU-Net baseline whose 95.4% accuracy and 35.49M parameters contextualize DDUNet's efficiency claim.","marker":"[11]"},{"why":"Provides the DeepLabV3+ baseline with comparable accuracy and far more parameters.","marker":"[36]"},{"why":"Supplies the inverted residual block used in DDUNet's decoders.","marker":"[32]"},{"why":"Provides the dilated-convolution context aggregation idea behind DMSC's multi-scale branches.","marker":"[33]"},{"why":"Supplies the pyramid context aggregation and short-cut connection design adapted into DMSC.","marker":"[34]"},{"why":"Defines the 9:1 SWINySEG split protocol used in the experiments and provides the MA-SegCloud baseline with the highest reported accuracy in the comparison table.","marker":"[38]"}],"fun_headline_variants":["Dynamic kernels make cloud segmentation 0.33M-parameter light","0.33M-parameter cloud segmenter hits 95.3% accuracy","U-Net with dynamic convs cuts parameters to 0.33M","Cloud segmentation goes ultra-light with dynamic U-Net","DDUNet: fast cloud segmentation with tiny model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison against prior models in Table I is fair, meaning the baseline numbers were obtained under the same 9:1 SWINySEG split and the same training protocol as DDUNet.","fun_headline_variants_meta":{"raw":{"variants":["Dynamic kernels make cloud segmentation 0.33M-parameter light","0.33M-parameter cloud segmenter hits 95.3% accuracy","U-Net with dynamic convs cuts parameters to 0.33M","Cloud segmentation goes ultra-light with dynamic U-Net","DDUNet: fast cloud segmentation with tiny model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000394,"raw_usage":{"total_tokens":2059,"prompt_tokens":926,"completion_tokens":1133,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":1043}},"tokens_in":542,"tokens_out":1133,"duration_ms":7686,"temperature":1.0,"reasoning_tokens":1043,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:19:56.048661+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released code on SWINySEG with the stated 9:1 split and the reported training schedule; the central claim fails if 95.3% accuracy and 0.884 MIoU do not reproduce, or if DeepLabV3+ and CloudU-Net, retrained under an identical protocol, match DDUNet's accuracy at a comparable parameter count.","supporting_citations":[{"cited_title":"U-net: Convo- lutional networks for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the U-Net encoder-decoder architecture that DDUNet builds on."},{"cited_title":"Cloud- segnet: A deep network for nychthemeron cloud image seg- mentation,","cited_arxiv_id":null,"evidence_quote":"Provides the CloudSegNet baseline with 0.005M parameters, the smallest model DDUNet compares against."},{"cited_title":"CloudU- Net: A Deep Convolutional Neural Network Architecture for Daytime and Nighttime Cloud Images’ Segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the CloudU-Net baseline whose 95.4% accuracy and 35.49M parameters contextualize DDUNet's efficiency claim."},{"cited_title":"Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully con- nected crfs,","cited_arxiv_id":null,"evidence_quote":"Provides the dilated-convolution context aggregation idea behind DMSC's multi-scale branches."},{"cited_title":"Pyramid scene parsing network,","cited_arxiv_id":null,"evidence_quote":"Supplies the pyramid context aggregation and short-cut connection design adapted into DMSC."},{"cited_title":"A Novel Ground-Based Cloud Image Segmentation Method Based on a Multibranch Asymmetric Convolution Module and Attention Mechanism,","cited_arxiv_id":null,"evidence_quote":"Defines the 9:1 SWINySEG split protocol used in the experiments and provides the MA-SegCloud baseline with the highest reported accuracy in the comparison table."}],"review_version":1}