{"id":"e906e4eb-76f8-4d80-a8f5-ce67d80ff208","arxiv_id":"2507.01721","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A convergent soft self-labeling framework with new Potts relaxations and a collision cross-entropy achieves state-of-the-art scribble-based semantic segmentation on standard architectures.","lead":"The paper trains segmentation networks from scribble annotations using soft pseudo-labels and a principled self-labeling loss. On standard architectures, the method matches or beats far more complex scribble-based systems, and in one configuration it even outperforms full pixel-level supervision.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The pseudo-label solver's near-global optimality is verified only for the convex PQ relaxation, not for the recommended non-convex PCD/PCCE terms; the main SOTA and full-supervision-overperformance claims therefore depend on an unverified solver assumption.","rationale":"The paper's central claim is that soft self-labeling with loss (8) and a general continuous pseudo-label solver consistently improves scribble-based training, outperforms specialized WSSS systems, and can outperform full pixel-precise supervision. The ADM derivation of loss (8) is algebraically correct, and the framework is principled. The most load-bearing assumption is that the pseudo-label subproblem is solved well enough for the recommended terms. Appendix A's only quantitative optimality check is for PQ, the convex quadratic relaxation, where a closed-form global solution exists; the recommended PCD, PCCE, and PLQ are non-convex, and the paper's own text acknowledges that the softmax reformulation makes the problem non-convex. Therefore the 99.2% soft Jaccard result does not transfer to the terms used in the headline results. The margins separating PCD from PQ and the scribble method from full supervision are small enough that solver local minima could change the conclusions. This is precisely the reader's weakest_assumption, and I agree with it. I do not see a separate mathematical flaw: the loss derivation, the relaxations, and the experimental comparisons are all coherent. The concern is that the central empirical claims are not robustly tied to a solver whose global behavior is verified for the recommended non-convex terms. This warrants a conditional verdict pending the proposed check, which is unchanged from the reader's verdict.","tokens_in":14376,"tokens_out":6509,"duration_ms":76972,"concrete_test":"Run the released code on Pascal VOC with the recommended HCCE+PCD configuration and the pseudo-label step initialized from (i) the paper's network-logit initialization, (ii) network logits plus Gaussian noise with std 0.3, and (iii) 5 random restarts per image on a 100-image subset; report the validation mIoU spread. In addition, on the same 100 images, compare PCD/PCCE solver outputs against a high-quality reference obtained by multi-start projected gradient descent on the original simplex-constrained problem, using soft Jaccard. If the PCD/PCCE validation mIoU varies across initializations by more than roughly 0.5 mIoU, or the soft Jaccard agreement with the reference is substantially below the 99.2% reported for PQ, the near-global-optimality assumption for the recommended relaxations is not supported and the headline claims are conditional on solver initialization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Appendix A states that solving for y at fixed sigma is a large-scale constrained convex problem, then replaces the simplex constraint with y = softmax(l), which the paper itself notes becomes non-convex. The only quantitative validation compares the solver against a closed-form global solution for the quadratic relaxation PQ and reports 99.2% soft Jaccard over 100 images. This check does not cover the terms actually recommended: PCD, PCCE, and PLQ are non-convex in y, so the 200-step gradient-descent solver initialized at network logits has no near-global optimality guarantee for them. If the solver is trapped in poor local minima for these terms, the pseudo-labels are not minimizers of loss (8), and the network is trained against suboptimal targets. The consequences are concrete: in Table 3 the margin of the best term PCD (71.22) over the convex PQ (71.05) is only 0.17 mIoU, and in Table 5 the margin over full supervision (77.7 vs 76.6) is 1.1 mIoU. A solver that finds slightly worse local minima for PCD could erase or invert both margins. This is an empirical gap, not a mathematical error, but the ablation claims in Table 3, Figure 5, and the SOTA comparison in Table 5 all inherit this uncertainty.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a soft self-labeling framework for weakly supervised semantic segmentation with scribble annotations. It derives an auxiliary loss (Eq. 8) by ADM splitting of a regularized loss (Eq. 4), introducing soft pseudo-labels y for unlabeled pixels and iterating between optimizing the network and the pseudo-labels. The authors systematically study several Potts relaxations (bilinear, quadratic, normalized quadratic, and log-based variants) and several cross-entropy-like terms (standard, reverse, and collision cross-entropy). A continuous gradient-descent solver for the pseudo-label subproblem is described in Appendix A. Experiments on Pascal VOC 2012, Cityscapes, and ADE20k report consistent gains over scribble-based baselines, state-of-the-art results among methods using standard architectures, and one configuration (DeepLabV3+, batch 12) where scribble supervision with the proposed loss (77.7 mIoU) outperforms training with full pixel masks (76.6 mIoU).","tokens_in":14659,"tokens_out":9341,"duration_ms":101320,"significance":"If the results hold, the paper offers a simple and reproducible recipe for scribble-supervised segmentation: standard architectures trained with loss (8) using collision cross-entropy and collision divergence outperform substantially more complex specialized WSSS systems and can approach or exceed full supervision. The algebraic identity H(sigma)+KL(sigma,y)=H(sigma,y) underlying Eq. (8) is clean, the evaluation is systematic across relaxations, cross-entropy terms, neighborhood systems, scribble lengths, backbones, and datasets, and the benchmark numbers are external and not fitted. The main uncertainty is the pseudo-label solver: its near-global optimality is verified only for the convex quadratic relaxation, while the recommended non-convex terms carry the headline claims. This is an empirical gap rather than a mathematical error, but it is load-bearing for the SOTA and full-supervision comparisons.","major_comments":[{"comment":"The pseudo-label solver is validated against global optima only for the convex quadratic relaxation PQ (99.2% soft Jaccard over 100 images), but the appendix itself notes that the softmax reparameterization makes the problem non-convex, and the recommended terms PCD, PCCE, and PLQ are non-convex in y. The main comparative claims rely on PCD (Table 5) and on the ranking in Table 3, where the margin between PCD and PQ is only 0.17 mIoU (71.22 vs 71.05); the full-supervision overperformance in Table 5 is 1.1 mIoU (77.7 vs 76.6). A solver that finds moderately worse local minima for PCD could erase or invert these margins. Please provide solver validation for the non-convex recommended terms (e.g., comparison with a stronger discrete or continuous solver on small problems, or a demonstration that the joint loss decreases monotonically in the alternating scheme), or explicitly qualify the SOTA and full-supervision claims.","section":"Appendix A, Table 3, Table 5"},{"comment":"The notation in Eq. (8) is internally inconsistent. The derivation from Eq. (4) uses the identity H(sigma_i)+KL(sigma_i,y_i)=H(sigma_i,y_i), so the H in Eq. (8) is the Shannon cross-entropy H(sigma,y), which is exactly the reverse cross-entropy HRCE of Eq. (10). Section 3.1 then selects HCCE (Eq. 11) as 'the cross-entropy term in (8)', but HCCE is not the Shannon cross-entropy and is not justified by the ADM derivation. The paper should either define H in Eq. (8) as a placeholder for a family of cross-entropy-style terms and state which convergence or monotonicity guarantees hold for each member, or clearly separate the derived loss from the heuristic HCCE variant.","section":"Section 2.2 and Eq. (8)"}],"minor_comments":[{"comment":"The symbol H(σ_i,y_i) is used both for the Shannon cross-entropy in the derivation and as a placeholder for HCE, HRCE, and HCCE later in the paper; please disambiguate the notation at first use.","section":"Eq. (8)"},{"comment":"The differences among PNQ, PQ, PCD, and PLQ at full scribbles are within 0.2 mIoU (71.05–71.22), so a single training run per configuration makes the reported ranking fragile; please report multiple seeds or state explicitly that the differences are not statistically significant.","section":"Table 3"},{"comment":"The sentence 'To solve y at given σ, it is a large-scale constrained convex problem' is accurate only for some terms (e.g., PQ with a convex unary); for the recommended non-convex terms it should be rephrased to avoid implying convexity.","section":"Appendix A"},{"comment":"The phrase 'guaranteed convergence' should be qualified to the specific loss and solver; the current wording may overstate the guarantee for the non-convex HCCE/PCD combination used in the headline experiments.","section":"Section 1.3 and Section 4"}],"recommendation":"major_revision","confidential_remarks":"The empirical claims are not circular: the benchmarks are external and no fitted constant is used to announce a prediction. The main risk is the solver-validation gap for the recommended non-convex Potts terms; if the authors can supply evidence that their gradient-descent solver reaches near-global solutions for PCD/PCCE, or at least that the alternating scheme decreases the joint loss, I would support acceptance. The paper would also benefit from a clearer statement of which variants are derived from the ADM splitting and which are heuristic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, you should know two things about this paper. It's a serious, well-derived framework for scribble-supervised segmentation that introduces soft self-labeling with several new Potts relaxations, and it reports numbers that beat more complex WSSS systems and sometimes full supervision. But the headline margins are thin, and the pseudo-label solver's near-global optimality is only validated for the convex relaxation, not for the recommended non-convex terms. The ADM derivation of loss (8) from loss (4) is clean; the H(sigma)+KL(sigma,y)=H(sigma,y) identity makes the auxiliary loss principled. The new normalized quadratic and log-based relaxations (PNQ, PCD, PCCE, PLQ) are motivated by concrete local-minima and vanishing-gradient examples; Figure 2's contours are informative. The collision cross-entropy (11) is a simple, principled way to avoid mimicking uncertain pseudo-labels, and the synthetic label-corruption experiment in Figure 3 supports it. The experiments are broad: three datasets, several backbones, and multiple baselines reproduced with public code. The main concern is the pseudo-label solver. Appendix A replaces the simplex constraint with y=softmax(l), notes it becomes non-convex, and uses 200 gradient steps initialized at network logits. The only quantitative check against global optima is for the convex quadratic PQ (99.2% soft Jaccard), not for the recommended PCD/PCCE. The margins in Table 3 (PCD vs PQ: 0.17 mIoU) and Table 5 (vs full supervision: about 1.1 mIoU) are small enough that a solver that lands in worse local minima for PCD could shift the rankings. This is an empirical gap, not a mathematical error; the authors do acknowledge the non-convexity, but validation is missing for the terms they recommend. Also, hyperparameters are tuned separately per variant and results are single-run without error bars, which weakens the ablations. Overall, the core derivation holds up and the paper deserves a serious referee. I'd ask reviewers to probe the pseudo-label solver's convergence on the non-convex terms, perhaps with multi-restart, and to request repeated runs for the key tables. It's a useful contribution for anyone working on weakly-supervised segmentation.","headline":"Solid soft self-labeling framework with clean derivations; empirical claims rest on a pseudo-label solver validated only for the convex relaxation, with thin margins.","tokens_in":15229,"tokens_out":3611,"would_cite":true,"duration_ms":35959,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Soft self-labeling with scribble annotations can beat full pixel-mask supervision.","keywords":["weakly-supervised semantic segmentation","scribble annotation","soft pseudo-labels","self-labeling","Potts model relaxation","collision cross-entropy","collision divergence","semantic segmentation"],"falsifier":"Run the recommended PCD/PCCE pseudo-label sub-problem on a small set of images with a global optimizer (an exhaustive multi-start search over the small neighborhood graph) and compare both the similarity of the pseudo-labels to the true optimum and the final network mIoU against the paper's 200-step gradient descent; if the global pseudo-labels do not reproduce the reported advantage over full supervision, the central claim depends on the solver rather than the loss.","tokens_in":14082,"feed_emoji":"🖼️","tokens_out":10732,"duration_ms":110012,"temperature":0.7,"pith_summary":"The paper claims that weakly supervised semantic segmentation from scribbles—ground-truth labels on only a few percent of pixels—can match or exceed full pixel-mask supervision when the network is trained with soft pseudo-labels instead of hard ones. The method augments the supervised loss on scribbles with a joint self-labeling loss over network predictions and auxiliary probability-distribution pseudo-labels, alternating network updates with a gradient-descent solver for the pseudo-labels. A systematic comparison of pairwise Potts relaxations and cross-entropy terms leads the authors to recommend a symmetric collision cross-entropy and a log-normalized quadratic pairwise term called collision divergence. On standard, unmodified architectures, the resulting loss outperforms specialized weakly-supervised systems and, in several settings, outperforms full supervision.","feed_headline":"Soft self-labeling beats full-mask training on scribble segmentation","feed_subtitle":"Training on scribbles, a soft pseudo-label loss can beat full pixel-mask supervision.","key_machinery":"The load-bearing object is the joint self-labeling loss (8), which couples network predictions $\\sigma_i$ with auxiliary soft pseudo-labels $y_i$ on the probability simplex: scribble pixels are fixed to ground truth, unlabeled pixels are free distributions. Alternating sub-problems—stochastic gradient descent on the network and a GPU gradient-descent solver on a softmax parameterization of the pseudo-labels—make the iteration convergent. The named innovations are collision cross-entropy $H_{\\mathrm{CCE}}(y,\\sigma) = -\\ln(\\sigma^\\top y)$, which maximizes the probability that predicted and pseudo-labeled classes agree without forcing the network to copy uncertainty, and collision divergence $P_{\\mathrm{CD}}$, the logarithm of a normalized quadratic Potts relaxation, which removes the flat-gradient and local-minimum problems of bilinear and plain quadratic forms.","core_discovery":"The central claim is that the bottleneck in scribble-supervised segmentation is the optimization of the unsupervised Potts loss, not the network architecture: replacing hard pseudo-labels with soft ones, and the standard cross-entropy target with a symmetric probability-of-collision term, produces a convergent self-labeling procedure that outperforms much more complex systems. The paper reports that on PASCAL VOC 2012, DeepLabV3+ trained with its loss and only scribbles reaches 77.7 mean intersection-over-union (mIoU) at batch size 12, versus 76.6 for the same architecture trained on full pixel masks; with a ViT-linear backbone it reaches 80.94 mIoU, above the best specialized system on the same backbone at 78.7 mIoU.","pith_inferences":["Editorial inference: the reported advantage over full supervision at batch size 12 may depend on how the full-supervision baselines were trained; a comparison with larger batches or stronger augmentation would test whether the gap is intrinsic.","Editorial inference: the paper checks its pseudo-label solver against global optima only for the convex quadratic relaxation, so the non-convex PCD/PCCE results leave open how much of the gain is solver-dependent.","Editorial inference: because the method makes no architectural assumptions, it should transfer to backbones beyond ResNet/MobileNet/ViT; testing on a newer convnet or transformer with stronger pretraining would extend the evidence.","Editorial inference: the robustness of collision cross-entropy to label uncertainty, shown in a corrupted-label classification experiment, suggests uses in noisy-label learning and knowledge distillation, though the paper does not claim those."],"forward_implications":["Scribble-supervised training with loss (8) becomes a drop-in replacement for full supervision on standard segmentation architectures, removing the need for specialized modules or multi-stage training.","The recommended combination of collision cross-entropy and collision divergence over a nearest-neighbor grid is the configuration to use; dense neighborhoods and non-normalized relaxations are systematically worse.","Because the pseudo-label solver can be run for 100 instead of 200 gradient steps at a cost of only 0.72 mIoU, the method offers a practical speed-accuracy trade-off.","Hard-pseudo-label self-labeling methods are limited by their inability to represent class uncertainty; any such system can in principle be upgraded to soft labels under the same joint-loss derivation."],"supporting_citations":[{"why":"Supplies the scribble-supervised segmentation setup and the hard-pseudo-label baseline that the paper improves on.","marker":"[28]"},{"why":"Introduces ADM splitting of regularized segmentation losses into network and pseudo-label sub-problems; the paper extends that idea from hard to soft labels.","marker":"[30]"},{"why":"Provides the prior hard self-labeling trust-region method whose results the paper compares against and generalizes.","marker":"[29]"},{"why":"Gives the quadratic Potts relaxation and its closed-form random-walker solution, used both as a baseline relaxation and as the check for the pseudo-label solver.","marker":"[19]"},{"why":"Supplies the bilinear tight relaxation and the convex quadratic relaxation of the Potts model that motivate the paper's normalized variants.","marker":"[35]"},{"why":"Defines the DenseCRF gradient-descent baseline and the dense-neighborhood setting that the paper re-examines and outperforms.","marker":"[38]"},{"why":"Contributes the bilateral-filtering trick for dense Potts gradients and the scribble pretraining initialization used in the experiments.","marker":"[37]"},{"why":"Provides soft proposal generation and the block-scribble supervision protocol used for the Cityscapes and ADE20k comparisons.","marker":"[27]"},{"why":"Is the specialized ViT-linear system whose reported mIoU the paper's standard-architecture method exceeds.","marker":"[43]"}],"fun_headline_variants":["Soft pseudo-labels beat full mask supervision on scribbles","Simple soft-label training beats complex segmentation systems","Scribble training with soft labels surpasses full supervision","A principled soft-label loss tops mask supervision and complex CRFs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the 200-step gradient-descent routine that estimates soft pseudo-labels lands close to the true optimum of a difficult, many-valley objective; if it gets stuck in a poor local valley, the reported gains would not transfer to a different solver or dataset.","fun_headline_variants_meta":{"raw":{"variants":["Soft pseudo-labels beat full mask supervision on scribbles","Simple soft-label training beats complex segmentation systems","Scribble training with soft labels surpasses full supervision","A principled soft-label loss tops mask supervision and complex CRFs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00019,"raw_usage":{"total_tokens":1310,"prompt_tokens":889,"completion_tokens":421,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":355}},"tokens_in":505,"tokens_out":421,"duration_ms":4940,"temperature":1.0,"reasoning_tokens":355,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:45:41.328489+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the recommended PCD/PCCE pseudo-label sub-problem on a small set of images with a global optimizer (an exhaustive multi-start search over the small neighborhood graph) and compare both the similarity of the pseudo-labels to the true optimum and the final network mIoU against the paper's 200-step gradient descent; if the global pseudo-labels do not reproduce the reported advantage over full supervision, the central claim depends on the solver rather than the loss.","supporting_citations":[{"cited_title":"Scribblesup: Scribble-supervised convolutional networks for semantic segmentation","cited_arxiv_id":null,"evidence_quote":"Supplies the scribble-supervised segmentation setup and the hard-pseudo-label baseline that the paper improves on."},{"cited_title":"Beyond gradient descent for regularized segmen- tation losses","cited_arxiv_id":null,"evidence_quote":"Introduces ADM splitting of regularized segmentation losses into network and pseudo-label sub-problems; the paper extends that idea from hard to soft labels."},{"cited_title":"Robust trust region for weakly supervised segmentation","cited_arxiv_id":null,"evidence_quote":"Provides the prior hard self-labeling trust-region method whose results the paper compares against and generalizes."},{"cited_title":"Random walks for image segmentation","cited_arxiv_id":null,"evidence_quote":"Gives the quadratic Potts relaxation and its closed-form random-walker solution, used both as a baseline relaxation and as the check for the pseudo-label solver."},{"cited_title":"Quadratic program- ming relaxations for metric labeling and Markov Random Field MAP estimation","cited_arxiv_id":null,"evidence_quote":"Supplies the bilinear tight relaxation and the convex quadratic relaxation of the Potts model that motivate the paper's normalized variants."},{"cited_title":"On reg- ularized losses for weakly-supervised cnn segmentation","cited_arxiv_id":null,"evidence_quote":"Defines the DenseCRF gradient-descent baseline and the dense-neighborhood setting that the paper re-examines and outperforms."},{"cited_title":"Normalized cut loss for weakly-supervised cnn segmentation","cited_arxiv_id":null,"evidence_quote":"Contributes the bilateral-filtering trick for dense Potts gradients and the scribble pretraining initialization used in the experiments."},{"cited_title":"Tree energy loss: Towards sparsely annotated semantic segmentation","cited_arxiv_id":null,"evidence_quote":"Provides soft proposal generation and the block-scribble supervision protocol used for the Cityscapes and ADE20k comparisons."},{"cited_title":"Modeling the Label Distributions for Weakly-Supervised Semantic Segmentation","cited_arxiv_id":"2403.13225","evidence_quote":"Is the specialized ViT-linear system whose reported mIoU the paper's standard-architecture method exceeds."}],"review_version":1}