{"id":"5f3feeb2-c141-4513-b59e-9029e32c5ae2","arxiv_id":"2501.01640","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A semi-supervised semantic segmentation method combining aleatoric uncertainty loss and energy-based loss with union-intersection pseudo-labels gives modest gains over the CPCL baseline on PASCAL VOC and Cityscapes, but not state-of-the-art results.","lead":"This paper adds uncertainty and energy-based loss terms to an existing semi-supervised segmentation framework with two pseudo-labeling branches. The reported gains over the baseline are small and the method trails the cited Unimatch approach on most benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed consistent gains over CPCL and SOTA are contradicted by the paper's own tables: Cityscapes 1/2 is below CPCL, and Unimatch beats DUEB ResNet-101 on PASCAL 1/16 and 1/8.","rationale":"The reader's weakest assumption was that the gains over CPCL may be within run-to-run variation because no error bars are reported and CPCL numbers are taken from the original paper. I agree with that concern, and I sharpen it: the inconsistency is visible in the paper's own tables, not merely a hypothetical seed effect. Table 3 shows DUEB ResNet-50 is worse than CPCL at Cityscapes 1/2 (77.58 vs 78.17), directly contradicting the claim of consistent gains. Table 4 shows DUEB ResNet-101 is below Unimatch ResNet-101 at PASCAL 1/16 and 1/8, directly contradicting the §4.2 SOTA claim for 'all partition protocols'. In addition, Table 2 shows the energy-only ablation is worse than the CE-only baseline at every PASCAL partition, so the attribution of the combined improvement to the energy loss is not supported. These are empirical and internal to the manuscript, not disagreements with community consensus. The proper fix is a controlled multi-seed comparison and a corrected SOTA statement; the idea is plausible and should not be rejected outright, so I keep the reader's conditional verdict. I mark agreement as partial because the reader emphasized external validity (no seeds, copied CPCL), while my primary concern is the direct internal contradiction between the claimed conclusions and the paper's own tables.","tokens_in":13724,"tokens_out":5781,"duration_ms":56469,"concrete_test":"Re-run DUEB and a re-implemented CPCL in one shared codebase on PASCAL VOC 1/4 and Cityscapes 1/2 and 1/16, using identical optimizer, learning-rate schedule, CutMix, and three random seeds, reporting mean and standard deviation of mIoU. If the DUEB−CPCL gap at Cityscapes 1/2 remains negative, or the 95% confidence interval overlaps zero at any partition, the claim of consistent improvement from the proposed losses fails. Independently, rebuild Table 4 including Unimatch ResNet-101 for all partition protocols and verify whether DUEB is actually best at 1/16 and 1/8.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central causal claim—that adding the aleatoric loss (Eq. 12) and energy loss (Eq. 16) to CPCL yields consistent mIoU gains—is not supported by the paper's own numbers. On Cityscapes, Table 3 gives DUEB ResNet-50 77.58 vs CPCL 78.17 at the 1/2 partition, i.e., the proposed losses hurt in the regime with the most labels. The SOTA statement in §4.2 is also internally contradicted by Table 4: Unimatch ResNet-101 reports 76.5/77.0 on PASCAL 1/16 and 1/8, while DUEB ResNet-101 scores 74.13/76.70. Furthermore, the only controlled baseline comparison is against CPCL numbers copied from [11], with no multiple seeds or error bars; most reported gains are 0.3–1.6 mIoU points, which is within plausible run-to-run variation. The ablation in Table 2 even shows the energy loss alone is below the CE-only baseline at every PASCAL partition (73.85 vs 74.58, 73.00 vs 73.74, 71.14 vs 71.66), so the contribution of the energy term is especially unclear. Taken together, the evidence as presented does not establish that the proposed losses cause the reported improvements.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DUEB, a semi-supervised semantic segmentation method built on the Conservative Progressive Collaborative Learning (CPCL) framework, adding an aleatoric (data) uncertainty loss and an energy-based loss to the standard intersection/union pseudo-label losses. The total loss in Eq. (17) combines detection loss, aleatoric loss, and energy loss, and the paper reports results on PASCAL VOC and Cityscapes at 1/2, 1/4, 1/8, and 1/16 partition protocols with ResNet-50 and ResNet-101 backbones. The central claim is that DUEB outperforms state-of-the-art methods, with consistent mIoU gains over CPCL. The paper also includes an ablation study and comparisons against several published baselines.","tokens_in":14040,"tokens_out":3599,"duration_ms":35058,"significance":"If the central claims were supported, the paper would make a useful contribution by showing that adding train-time aleatoric uncertainty and a simple energy-based term to pseudo-label supervision improves semi-supervised semantic segmentation. The method is built on a strong existing baseline (CPCL), and the authors provide a code link and experiments across two standard benchmarks at multiple partition ratios, which is helpful for reproducibility. However, the significance is substantially diminished by internal numerical contradictions: several of the paper's own tables contradict the claim of state-of-the-art performance, and the ablation table is unreadable. The absence of error bars and the reliance on baseline numbers copied from prior work also make the reported gains hard to interpret. The conceptual framing of the energy loss as 'generative modeling' is not backed by the equations, which describe a log-sum-exp confidence maximization.","major_comments":[{"comment":"The claim that 'DUEB with ResNet-101 backbone outperforms the state-of-the-art methods for all partition protocols' is directly contradicted by the reported numbers. On PASCAL VOC, Unimatch ResNet-101 achieves 76.5 at 1/16 and 77.0 at 1/8, while DUEB ResNet-101 achieves 74.13 and 76.70, respectively; on Cityscapes at 1/2, Unimatch ResNet-101 achieves 79.5 versus DUEB's 79.47. The sentence also incorrectly cites Tables 1 and 3, which contain only ResNet-50 results, rather than Tables 4 and 5. This overstatement must be corrected and the claim qualified to the configurations where the comparison actually holds.","section":"§4.2, Table 4 and Table 5"},{"comment":"The ablation table is not interpretable as printed: the checkmarks and numeric values are arranged without a clear row/column layout, and some cells (e.g., the PASCAL VOC 1/4 entry with both losses checked) appear missing. The reader cannot verify the stated conclusion that the combination of uncertainty and energy losses outperforms either component alone. The table must be redrawn with explicit rows for each loss configuration and columns for each dataset and partition.","section":"Table 2"},{"comment":"The paper reports no error bars, no multiple-seed runs, and the CPCL baseline numbers are taken from the original CPCL paper rather than re-trained under the same pipeline. Several reported gains over CPCL are below one mIoU point (e.g., PASCAL 1/2: 75.94 vs 75.30; Cityscapes 1/4: 77.85 vs 76.98), which is within plausible run-to-run variation in this setting. Without repeated trials, the causal claim that the proposed losses are responsible for the improvement is not supported by the evidence presented.","section":"§4.2, Tables 1–5"},{"comment":"The aleatoric loss L_ale is under-specified. The expression '(-ELU * diff) * lu' is unclear, especially with regard to how the exponential linear unit is applied to a potentially negative 'diff' and how the resulting value multiplies the cross-entropy term; the variance term 'e^{σ2} - 1' also has an ambiguous role in the total. Since this loss is one of the two central contributions, the equation must be defined unambiguously, including all indices, the Monte-Carlo average over T, and the exact inputs (logits, variance, and pseudo-label) used for the conservative and progressive branches.","section":"§3.3, Eq. (12)"},{"comment":"The energy loss L_e = LogSumExp(fθ(x)|y) is the negative of the energy function E_θ(x) defined in Eq. (15). Minimizing this loss therefore maximizes the log-sum-exp of logits, which is essentially a confidence-maximization term; the paper does not train the partition function or sample from the implicit energy distribution over inputs. Consequently, the claim that the loss provides 'generative modeling' and learns the joint distribution p(x, y) is not supported by the equations. The authors should either provide the actual EBM training mechanism (e.g., contrastive divergence, noise-contrastive estimation) or substantially moderate the claimed role of the energy term.","section":"§3.4, Eq. (15)–(16)"}],"minor_comments":[{"comment":"There are several typographical and grammatical issues: 'availaible', 'pseudolabel' vs 'pseudo label' inconsistently, 'enormous pixel-label annotated data', and the sentence fragment 'The other constraint with the SS segmentation methods is the discriminative and deterministic framework, which fails to capture the generative distribution.' These should be cleaned up.","section":"Abstract and §1"},{"comment":"The sentence 'The best performance is achieved by the vehicle group, which is 19.77% over the supervised CPCL baseline and 4.69% over CPCL' lacks a clear baseline reference and appears to conflate percentage-point differences with percentages; the numbers should be rechecked and stated consistently.","section":"§4.2"},{"comment":"Eq. (17) refers to 'the loss using the union and intersection label as pesdueo label defined in Eq.', with the equation number missing. The reference should be completed.","section":"§3.5, Eq. (17)"},{"comment":"The class-group abbreviations in the table headers are defined in the caption but the captions themselves are overly long and some definitions are grammatically incomplete; additionally, the table footnotes mix 'ANIMAL' and 'Animal' capitalization inconsistently.","section":"Tables 1 and 3"},{"comment":"The contribution bullet 'Enhance the SS prediction by energy-based loss to incorporate generative modeling using the discriminative function' is vague and should be made more specific after the energy-loss issue in the major comments is resolved.","section":"§1, Introduction"}],"recommendation":"major_revision","confidential_remarks":"The numerical contradictions between the claims and the tables are the main obstacle; the authors need to either retrain with proper comparisons and error bars or substantially weaken the claims. The garbled ablation table also needs to be fixed. I would not recommend rejection at this stage, as the losses themselves are not circular and the method might yield modest gains under a properly controlled comparison, but the manuscript in its current form does not support its central assertions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper grafts two known losses—aleatoric variance loss and energy-based LogSumExp loss—onto CPCL's semi-supervised segmentation framework. The only new element is the joint application. The method is clearly described, but the central claim of consistent gains over CPCL and state-of-the-art is contradicted by the paper's own tables.\n\nWhat the paper does well: it identifies a reasonable research direction and evaluates it on two benchmarks with several partitions and backbones. The PASCAL results show a consistent, if small, improvement over CPCL, and the uncertainty loss alone does help in the ablation. This is a plausible plug-in worth investigating.\n\nBut the central claim does not hold up. On Cityscapes at 1/2 partition, DUEB ResNet-50 gets 77.58 vs CPCL's 78.17, so the proposed losses hurt in the most labeled regime. The Section 4.2 statement that DUEB with ResNet-101 'outperforms the state-of-the-art methods for all partition protocols' is false by the paper's own Tables 4 and 5. Unimatch beats DUEB on most configurations: on Cityscapes, Unimatch is better at every partition (e.g., 76.6 vs 74.16 at 1/16), and on PASCAL it wins at 1/16 and 1/8.\n\nThe only controlled comparison is against CPCL numbers copied from the original paper, with no retraining, no error bars, no multiple seeds. Most gains are 0.5–1.5 mIoU points, which is within run-to-run variation. The ablation table is garbled, but as far as I can read it, the energy loss alone is worse than the cross-entropy-only baseline on every PASCAL partition. That makes the contribution of the energy term especially unclear.\n\nWho this is for: someone working on pseudo-label refinement in semi-supervised segmentation, especially if they want a simple baseline combining uncertainty and energy losses. A serious referee should engage with it, but the paper needs major revision: re-run with multiple seeds, report error bars, fix the SOTA claim, and clean up the ablation.\n\nRecommendation: send it to peer review, not desk reject, because the idea is legitimate and the experiments are extensive, but demand the above corrections before acceptance.","headline":"A straightforward combination of known losses whose claimed consistent gains are contradicted by its own tables; the method may work but the evidence as presented is not convincing.","tokens_in":14570,"tokens_out":4019,"would_cite":false,"duration_ms":34789,"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":"Semi-supervised segmentation improves by adding a per-pixel data-uncertainty loss and an energy-based loss to CPCL's union/intersection pseudo-label loss, with mIoU gains reported on PASCAL VOC and Cityscapes.","keywords":["semi-supervised semantic segmentation","aleatoric uncertainty","energy-based model","pseudo-label refinement","intersection-union pseudo labels","DeepLabv3+","data uncertainty"],"falsifier":"Retrain CPCL and DUEB from the same initializations with at least five seeds on PASCAL VOC 1/16 and Cityscapes 1/16, using the exact same augmentation, learning schedule, and code base; if the mean DUEB mIoU does not exceed retrained CPCL by more than the seed standard deviation, the central claim fails. Alternatively, an ablation that removes only the two new terms while keeping the variance layer should show a drop larger than the seed noise.","tokens_in":13539,"feed_emoji":"🎯","tokens_out":6109,"duration_ms":49586,"temperature":0.7,"pith_summary":"Labeling every pixel in an image is expensive, so semi-supervised segmentation tries to learn from a few labeled images plus many unlabeled ones. This paper proposes DUEB, a training loss that adds two terms to the CPCL baseline: a data-uncertainty (aleatoric) loss that measures how noisy the network's own per-pixel logits are, and an energy-based loss that treats the classifier as a generative energy model. The authors claim that combining these losses with CPCL's intersection/union pseudo-labels improves mean intersection-over-union (mIoU) on PASCAL VOC and Cityscapes at every labeled-data partition for ResNet-50 except one, and for ResNet-101 at all partitions. The largest relative gains occur when labeled data is scarcest, at the 1/16 partition.","feed_headline":"Uncertainty and energy losses lift semi-supervised segmentation","feed_subtitle":"Gains over CPCL show up on PASCAL VOC and Cityscapes, largest when labels are few.","key_machinery":"The paper's central object is the total loss in Eq. 17, which adds two terms to the CPCL detection loss $L_{det}$. First, the aleatoric uncertainty loss (Eq. 12) uses a variance layer appended to DeepLabv3+ that outputs a per-pixel variance $\\sigma_x$; it computes Monte-Carlo (T=10) distorted logits and penalizes the difference between undistorted and distorted cross-entropy plus an exponential variance penalty. Second, the energy loss (Eq. 16) is the LogSumExp over class logits, which, following the identity that a classifier can be reinterpreted as an energy-based model, encourages the model to assign high density to data. Each loss is applied on the conservative branch with intersection pseudo-labels, on the progressive branch with union pseudo-labels, and (for the energy loss) on ground-truth supervision; the two branches and their union/intersection pseudo-label generation come directly from CPCL.","core_discovery":"The central claim is that the total loss $L_{total} = L_{det} + \\gamma_{ale}(L^{c}_{ale}+L^{p}_{ale}) + \\gamma_e(L^{c}_e + L^{p}_e)$ (Eq. 17) yields better semi-supervised segmentation than the CPCL baseline, with both hyperparameters set to 1. The aleatoric loss (Eq. 12) makes the network output a per-pixel variance $\\sigma_x$ alongside logits, then computes cross-entropy on original logits and on logits distorted by Gaussian noise $\\epsilon_t \\sim N(0,\\sigma_x)$, penalizing the difference plus an exponential variance term $e^{\\sigma^2}-1$. The energy loss (Eq. 16) maximizes the LogSumExp of the logits, which is equivalent to minimizing the free energy of a joint energy-based model. On PASCAL VOC with ResNet-50, DUEB reaches 75.94 vs 75.30 mIoU at 1/2 labeled data and 72.41 vs 71.66 at 1/16; on Cityscapes it reaches 77.85 vs 76.98 at 1/4, 76.16 vs 74.60 at 1/8, and 72.38 vs 69.92 at 1/16, while at 1/2 it drops slightly (77.58 vs 78.17). With the ResNet-101 backbone the paper reports outperforming all compared state-of-the-art methods at every partition.","pith_inferences":["The reported improvements over CPCL are mostly under 1 mIoU point on PASCAL VOC and are compared against CPCL numbers taken from the original paper rather than a re-trained baseline; without multiple seeds or error bars, part of the gain could be training noise.","A natural test is to run DUEB against CPCL re-trained in the identical pipeline with several seeds; if the gap disappears or reverses, the claim of consistent improvement would weaken.","The aleatoric loss can be interpreted as a form of logit regularization that might transfer to other dense prediction tasks such as depth estimation or detection.","The energy loss's LogSumExp term is essentially a softmax-denominator regularizer; it would be worth testing whether simpler penalties (e.g., entropy) reproduce the same gains, which would suggest the mechanism is not specifically generative."],"forward_implications":["If correct, existing semi-supervised segmentation networks can be upgraded by adding a variance output layer and the two loss terms, with no new data or architectural change beyond that.","The method gives the largest mIoU gains when labeled data is scarce, making it relevant to annotation-limited domains such as medical imaging or autonomous driving.","The energy loss provides a generative interpretation for a discriminative segmenter, which could improve robustness and calibration beyond mIoU.","The gains at the 1/2 Cityscapes partition are negative relative to CPCL, so the benefit is not uniform across all label regimes.","The proposed framework is presented as generic and applicable to other semi-supervised segmentation networks."],"supporting_citations":[{"why":"Defines DeepLabv3+, the backbone architecture to which the variance layer is added for per-pixel logits and uncertainty.","marker":"[4]"},{"why":"Supplies the conservative-progressive collaborative learning framework, the intersection/union pseudo-labels, and the CPCL baseline mIoU numbers that DUEB compares against.","marker":"[11]"},{"why":"Establishes the claim that a discriminative classifier can be viewed as an energy-based model over the joint distribution, providing the theoretical basis for the LogSumExp energy loss in Eq. 16.","marker":"[13]"},{"why":"Introduces the heteroscedastic aleatoric uncertainty formulation with Gaussian logit noise that Eq. 12 builds upon.","marker":"[19]"},{"why":"Motivates using joint energy-based models in semi-supervised settings, which supports the addition of the energy loss for uncertainty calibration.","marker":"[45]"}],"fun_headline_variants":["Aleatoric and energy losses boost semi-supervised segmentation","Uncertainty plus energy loss improves pseudo-label segmentation","Semi-supervised segmentation gains from uncertainty and energy modeling","Data uncertainty and energy loss sharpen pseudo-supervised segmentation","Aleatoric uncertainty and energy loss advance SS segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper attributes the mIoU gains to the new losses, but because the CPCL baseline numbers are taken from the original publication rather than retrained in the same pipeline and no multiple-seed runs or error bars are reported, gains under 1 mIoU point could be run-to-run variation.","fun_headline_variants_meta":{"raw":{"variants":["Aleatoric and energy losses boost semi-supervised segmentation","Uncertainty plus energy loss improves pseudo-label segmentation","Semi-supervised segmentation gains from uncertainty and energy modeling","Data uncertainty and energy loss sharpen pseudo-supervised segmentation","Aleatoric uncertainty and energy loss advance SS segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000718,"raw_usage":{"total_tokens":3262,"prompt_tokens":1018,"completion_tokens":2244,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":634,"completion_tokens_details":{"reasoning_tokens":2166}},"tokens_in":634,"tokens_out":2244,"duration_ms":16155,"temperature":1.0,"reasoning_tokens":2166,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:23:05.962041+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain CPCL and DUEB from the same initializations with at least five seeds on PASCAL VOC 1/16 and Cityscapes 1/16, using the exact same augmentation, learning schedule, and code base; if the mean DUEB mIoU does not exceed retrained CPCL by more than the seed standard deviation, the central claim fails. Alternatively, an ablation that removes only the two new terms while keeping the variance layer should show a drop larger than the seed noise.","supporting_citations":[{"cited_title":"Encoder-decoder with atrous separable convolution for semantic image segmentation","cited_arxiv_id":null,"evidence_quote":"Defines DeepLabv3+, the backbone architecture to which the variance layer is added for per-pixel logits and uncertainty."},{"cited_title":"Conservative-progressive collab- orative learning for semi-supervised semantic segmentation","cited_arxiv_id":null,"evidence_quote":"Supplies the conservative-progressive collaborative learning framework, the intersection/union pseudo-labels, and the CPCL baseline mIoU numbers that DUEB compares against."},{"cited_title":"Your classifier is secretly an energy based model and you should treat it like one","cited_arxiv_id":null,"evidence_quote":"Establishes the claim that a discriminative classifier can be viewed as an energy-based model over the joint distribution, providing the theoretical basis for the LogSumExp energy loss in Eq. 16."},{"cited_title":"What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017","cited_arxiv_id":null,"evidence_quote":"Introduces the heteroscedastic aleatoric uncertainty formulation with Gaussian logit noise that Eq. 12 builds upon."},{"cited_title":"Joint energy-based models for semi-supervised classifica- tion","cited_arxiv_id":null,"evidence_quote":"Motivates using joint energy-based models in semi-supervised settings, which supports the addition of the energy loss for uncertainty calibration."}],"review_version":1}