{"id":"ae6eac33-e186-4007-b0a7-f2c57158e039","arxiv_id":"2506.13043","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ViewPCL uses Wasserstein distance between cross-view point cloud distributions as an uncertainty score, and it reports higher mIoU than ViewAL on SceneNet-RGBD.","lead":"ViewPCL is a new active learning method for multi-view segmentation that scores image regions by geometric disagreement between camera views, measured as a Wasserstein distance between class-conditional point cloud distributions. The authors report that selecting the most inconsistent superpixels for labeling improves data efficiency over the ViewAL baseline on the SceneNet-RGBD indoor dataset.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline gain over ViewAL is not shown to be statistically reliable: 5 seeds, one dataset, one baseline, no error bars, and unequal run lengths; 'clearly outperforms' is under-supported.","rationale":"The reader's weakest_assumption—that the top-1-class Wasserstein score may discard informative lower-ranked classes—is a legitimate algorithmic concern, and the lack of ablation there is real. However, the headline claim is empirical, so the most load-bearing condition is that the measured improvement over ViewAL is statistically meaningful. The paper provides no confidence information, and the description of experiments in Section 5 introduces an unexplained discrepancy in the number of rounds run per seed. This makes the central claim vulnerable even before asking whether the score is optimal. My recommendation therefore leaves the reader's CONDITIONAL verdict unchanged: the method is plausible and clearly described, but the key empirical assertion needs a stronger statistical demonstration. A paired significance check or bootstrapped confidence interval on the per-seed differences would settle whether the current evidence supports 'clearly outperforms' or whether the claim should be softened to 'does not degrade on this benchmark.'","tokens_in":90,"tokens_out":3798,"duration_ms":57073,"concrete_test":"Run the released code for both ViewPCL and ViewAL on the same 5 seeds for exactly 4 active-learning rounds, record per-round mIoU differences per seed, and compute a paired bootstrap or Wilcoxon 95% confidence interval on the difference at each round. Also check whether restricting the average to the 3 seeds that ran 7 rounds changes the ordering. If any round's interval includes zero, or the ordering flips under the 3-seed restriction, the 'clearly outperforms' claim is not supported by the current experiment.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is the empirical sentence in Section 5: 'Our proposed method clearly outperforms ViewAL on the SceneNet-RGBD dataset.' For that claim to hold, the observed mIoU advantage of ViewPCL over ViewAL must be larger than run-to-run variability from random seeds, initialization, and stochastic training. The reported evidence does not establish this: Figure 1 shows raw and averaged curves over 5 seeds but reports no error bars, variance, or significance test; only one synthetic dataset and one baseline are used; and Section 5 itself notes an unexplained split, where 3 experiments ran 7 active-learning rounds while the other 2 ran only 4. Averaging the first 4 rounds across all 5 runs mixes two protocols, and without per-seed paired differences it is impossible to tell whether the advantage is systematic or falls within noise. This is load-bearing because if the difference is within noise, the central claim reduces to an unquantified single-benchmark observation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ViewPCL, an active learning method for multi-view semantic segmentation. The method introduces a superpixel-level uncertainty score based on the Wasserstein distance between point cloud distributions induced by class probability maps from different views, aggregated over overlap regions. The selection policy combines this new 'viewpcl inconsistency score' with the view divergence score from ViewAL. The method is evaluated on SceneNet-RGBD against ViewAL using DeepLabv3+ with MobileNetV2, with 5 random seeds, and the authors claim that ViewPCL 'clearly outperforms' ViewAL in terms of mIoU at the same labeling budget.","tokens_in":8462,"tokens_out":3151,"duration_ms":35003,"significance":"If the empirical claim is established, ViewPCL would provide a new, geometrically interpretable uncertainty signal for multi-view active learning, with a clearly formulated score and released source code, which would be a useful contribution to the active learning literature. The method's derivation is internally consistent and does not appear circular: the score is defined a priori from model predictions and is not fitted to the outcome. However, the current evidence for the central claim is statistically weak: no error bars, no significance tests, a single baseline, a single dataset, and unexplained unequal experiment lengths. The contribution is therefore promising but not yet convincingly demonstrated.","major_comments":[{"comment":"The central claim that ViewPCL 'clearly outperforms ViewAL' is not supported by the reported statistics. Figure 1 shows raw and averaged mIoU curves over 5 seeds but provides no error bars, variance, confidence intervals, or significance tests. With only 5 runs, the observed advantage could lie within run-to-run variability. The authors should report per-seed paired differences between ViewPCL and ViewAL, and perform a paired statistical test (e.g., Wilcoxon signed-rank or permutation test) on the mIoU values at each round, or provide confidence intervals. Without this, the headline claim is not quantitatively grounded.","section":"Section 5, Figure 1"},{"comment":"The experiments have unequal lengths: 3 runs were executed for 7 active learning rounds and 2 for 4 rounds. The text does not explain this discrepancy, yet Figure 1b averages over all 5 runs for the first 4 rounds, mixing two different protocols. This makes the average curve potentially misleading. The authors should either run all experiments for the same number of rounds or analyze the two subgroups separately, and should state the reason for the unequal protocol.","section":"Section 5"},{"comment":"The evaluation uses only one baseline (ViewAL) and one dataset (SceneNet-RGBD). To support the general claim of data efficiency, the comparison should include at least a random-sampling baseline and a standard uncertainty-sampling baseline (e.g., entropy or margin), and ideally a second dataset. Furthermore, the ViewPCL selection policy in Section 3.4 combines the new viewpcl inconsistency score with ViewAL's view divergence score; without an ablation that removes one component, the observed improvement cannot be attributed to the proposed score rather than to the particular hybrid policy.","section":"Section 4.5 and Section 5"},{"comment":"The score in Eq. (s(D)) uses only the most probable class per view (ĉ1 and ĉ2) and a uniform selection distribution q. The authors acknowledge the flexibility of q but do not ablate these choices. If the model's top-1 class is wrong in both views, or if informative uncertainty resides in lower-ranked classes, the score may miss valuable regions. An ablation varying the number of classes considered or the selection distribution q would strengthen the claim that the proposed score captures meaningful cross-view uncertainty.","section":"Section 3.3.2"}],"minor_comments":[{"comment":"The superpixel algorithm is referred to as 'SEEDs' but the standard name is 'SEEDS'. Please correct the capitalization.","section":"Section 4.2"},{"comment":"References [12] and [13] are the same paper (Gal et al., ICML 2017). Please merge or remove the duplicate.","section":"References"},{"comment":"The notation W in the score definition is not fully specified; the text later mentions that the Wasserstein distance is approximated by sliced Wasserstein, but it would be clearer to state this at the point of definition.","section":"Section 3.3.2"},{"comment":"The caption of Figure 1 says '4 active learning rounds' but the text explains that only 2 experiments ran exactly 4 rounds while the other 3 ran 7 rounds. Please clarify whether the figure shows only the first 4 rounds of the longer runs or only the 2 short runs, and make the caption consistent.","section":"Section 5, Figure 1"},{"comment":"The fallback transport plan that moves µq,c1 to the boundary of the bounding box is heuristic. It would be helpful to report how often this fallback is triggered in practice, since the main text claims it occurs only in 'rare cases'.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The paper's main weakness is empirical: the central comparison to ViewAL lacks statistical rigor and baseline breadth. The method itself is clearly described and the code is released, which are strong positives. I recommend major revision rather than rejection because the flaw is in the evidence supporting the claim, not in the core idea, and it is fixable within the scope of the manuscript by adding error bars, paired tests, and at least one additional baseline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new idea here is ViewPCL's inconsistency score: instead of aggregating pixel-wise view disagreement as ViewAL does, it builds class-conditional point cloud distributions over a superpixel and measures their Wasserstein distance across views. That is genuinely new, it is well motivated, and the derivation in Section 3 is internally consistent. The sliced-Wasserstein implementation is sensible, the selection policy is a simple and transparent modification of ViewAL, and the source code is public. On the substance, this is a legitimate contribution to the multi-view active learning niche, not a broad breakthrough.\n\nThe central problem is the evidence for the headline claim. Section 5 says the method 'clearly outperforms ViewAL,' but the support is one synthetic dataset, one baseline, five seeds, and no error bars, confidence intervals, or significance test. The stress-test concern holds on reading the paper: without per-seed paired differences, the observed mIoU gap could easily be within run-to-run noise. The unexplained split—3 experiments ran 7 rounds, 2 ran 4—does not necessarily invalidate the first four rounds, since the early rounds may follow the same protocol, but it should be justified and reported cleanly. The absence of random-sampling or entropy-sampling baselines also makes it hard to calibrate how much of the gain is the new score versus the superpixel policy itself.\n\nTwo smaller soft spots. The score only compares the most probable class from each view, and there is no ablation on whether the top-1 assumption misses informative uncertainty in lower-ranked classes. This is a modeling choice, not a fatal flaw, but it deserves a test. The hyperparameters (overlap threshold, weight threshold, number of MC dropout runs, number of slices) appear hand-chosen; that is normal, but a sensitivity sweep would strengthen the paper.\n\nTo be fair, the method is not circular: the score is defined from model predictions before any fitting to outcomes, and the evaluation is a post-hoc comparison. The paper also acknowledges the irregular run lengths, which suggests the authors are not hiding the issue. The core contribution is real and testable, but the empirical claim needs stronger support before I would trust the 'clearly outperforms' language.\n\nRecommendation: send it to peer review. A serious referee should ask for error bars or paired comparisons, at least one more baseline and one more dataset, and an ablation on the number of classes used in the score. With those additions, this could be a solid methods paper.","headline":"A clean, geometrically motivated Wasserstein inconsistency score for multi-view active learning, with an empirical claim that currently outruns the evidence: one dataset, one baseline, no error bars.","tokens_in":8940,"tokens_out":2128,"would_cite":false,"duration_ms":29463,"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":"ViewPCL's Wasserstein-based inconsistency score selects superpixels whose cross-view point cloud disagreement is largest, and this policy beats ViewAL on SceneNet-RGBD.","keywords":["active learning","multi-view segmentation","semantic segmentation","uncertainty sampling","Wasserstein distance","superpixel selection","view consistency","point cloud distributions"],"falsifier":"An ablation on SceneNet-RGBD that computes the score over all classes instead of the top-1 class and shows no mIoU gain (or a loss) would undercut the claim that top-1 geometric disagreement is the informative signal. Alternatively, a comparison on a real indoor RGB-D dataset with noisy depth where ViewPCL fails to beat ViewAL would bound the claim to synthetic data.","tokens_in":8027,"feed_emoji":"🎯","tokens_out":3191,"duration_ms":29950,"temperature":0.7,"pith_summary":"ViewPCL proposes an active learning method for multi-view semantic segmentation that decides which superpixels an annotator should label. Its new viewpcl inconsistency score measures, with a Wasserstein distance, how much two views of the same surface disagree on where a predicted object class sits in 3D. The paper argues that this geometric disagreement signal is more informative than the pixel-wise viewpoint entropy used by ViewAL, and reports that ViewPCL achieves higher mean IoU than ViewAL on SceneNet-RGBD at the same labeling budget. A sympathetic reader should care because the method promises better data efficiency for segmentation models when annotations are scarce.","feed_headline":"Point-cloud disagreement picks pixels worth labeling","feed_subtitle":"A Wasserstein-based score beats ViewAL on SceneNet-RGBD at the same annotation budget.","key_machinery":"The load-bearing object is the viewpcl inconsistency score s(R). For a superpixel R, the method computes overlapping subregions D where another image's class probability map cross-projects onto R, forms point cloud distributions $\\mu^{q,\\hat{c}_1}_1$ and $\\mu^{q,\\hat{c}_1}_2$ for the most probable class $\\hat{c}_1$ of the reference view, and sets $s(D)$ as the average of two Wasserstein distances between these distributions (one for $\\hat{c}_1$, one for the other view's most probable class $\\hat{c}_2$). The superpixel score $s(R)$ is the size-weighted average of $s(D)$ over overlap subregions. The selection policy ranks superpixels by this score, then, like ViewAL, picks the cross-projected superpixel with highest view divergence to annotate.","core_discovery":"The paper's central claim is empirical: with the same number of labeled superpixels, the ViewPCL acquisition policy produces a segmentation model with higher mIoU on SceneNet-RGBD than the ViewAL policy. The score driving the policy is the viewpcl inconsistency score, which for each superpixel averages Wasserstein distances between point cloud distributions induced by class probability maps from the reference view and from cross-projected views. The method treats disagreement about the geometry of the most probable class as the uncertainty to resolve. The paper also claims this score is interpretable: it measures how badly two views disagree on the configuration of an object within a region.","pith_inferences":["The top-1 class restriction could be relaxed to a top-k or all-class Wasserstein sum; a natural test is whether multi-class disagreement further improves acquisition on scenes with many small objects.","One could replace the uniform selection distribution q with an uncertainty-weighted distribution, a direction the authors themselves flag as future work; this might sharpen the score in early rounds.","The method's reliance on accurate pose and depth means its gains may shrink on real-world data with noisy depth; testing on real indoor RGB-D scans would delineate the regime where the geometric score pays off.","Cross-view consistency is a strong prior; on datasets where the same surface point is rarely seen from multiple views, the score would carry little signal and random sampling might match it."],"forward_implications":["If the claim holds, active learning for multi-view segmentation can be driven by cross-view geometric disagreement rather than pixel-wise entropy.","The same score could be applied to other dense prediction tasks with pose and depth, such as multi-view object detection or depth completion.","Because the score is computed from predicted class probability maps, it requires no extra labels and can be used in any round of active learning.","The reported improvement on SceneNet-RGBD suggests synthetic indoor datasets are a suitable testbed for view-consistency acquisition policies.","The method inherits ViewAL's superpixel selection and thus keeps annotation effort at the superpixel level."],"supporting_citations":[{"why":"Baseline method and source of the dataset subsets, preprocessing, MC dropout protocol, and the selection policy that ViewPCL modifies.","marker":"[33]"},{"why":"Provides the SceneNet-RGBD synthetic indoor trajectory dataset used in the evaluations.","marker":"[21]"},{"why":"Companion paper for SceneNet-RGBD; referenced for the dataset and its relevance to indoor segmentation.","marker":"[22]"},{"why":"Supplies the Wasserstein distance formalism the inconsistency score is built on.","marker":"[28]"},{"why":"Provides the Slice Wasserstein Distance implementation used to compute point cloud dissimilarities.","marker":"[10]"},{"why":"SEEDs algorithm used to generate the superpixels that define the annotation regions.","marker":"[4]"}],"fun_headline_variants":["Wasserstein view gaps pick pixels to label","Point-cloud disagreement outperforms ViewAL","Cross-view point clouds score label value","View geometry inconsistency beats ViewAL","Active segmentation via Wasserstein views"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The score sums disagreement only for the single most probable class from each view, so if the most informative uncertainty lies in lower-ranked classes, the acquisition signal may miss the most valuable superpixels.","fun_headline_variants_meta":{"raw":{"variants":["Wasserstein view gaps pick pixels to label","Point-cloud disagreement outperforms ViewAL","Cross-view point clouds score label value","View geometry inconsistency beats ViewAL","Active segmentation via Wasserstein views"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000182,"raw_usage":{"total_tokens":1195,"prompt_tokens":715,"completion_tokens":480,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":331,"completion_tokens_details":{"reasoning_tokens":419}},"tokens_in":331,"tokens_out":480,"duration_ms":5677,"temperature":1.0,"reasoning_tokens":419,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:36:38.396566+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"An ablation on SceneNet-RGBD that computes the score over all classes instead of the top-1 class and shows no mIoU gain (or a loss) would undercut the claim that top-1 geometric disagreement is the informative signal. Alternatively, a comparison on a real indoor RGB-D dataset with noisy depth where ViewPCL fails to beat ViewAL would bound the claim to synthetic data.","supporting_citations":[{"cited_title":"Viewal: Active learning with view- point entropy for semantic segmentation","cited_arxiv_id":null,"evidence_quote":"Baseline method and source of the dataset subsets, preprocessing, MC dropout protocol, and the selection policy that ViewPCL modifies."},{"cited_title":"Scenenet rgb-d: 5m photorealistic images of synthetic indoor trajec- tories with ground truth","cited_arxiv_id":null,"evidence_quote":"Provides the SceneNet-RGBD synthetic indoor trajectory dataset used in the evaluations."},{"cited_title":"Scenenet rgb-d: Can 5m synthetic images beat generic imagenet pre-training on indoor segmentation?","cited_arxiv_id":null,"evidence_quote":"Companion paper for SceneNet-RGBD; referenced for the dataset and its relevance to indoor segmentation."},{"cited_title":"Optimal transport for applied mathematicians","cited_arxiv_id":null,"evidence_quote":"Supplies the Wasserstein distance formalism the inconsistency score is built on."},{"cited_title":"Pot: Python optimal transport","cited_arxiv_id":null,"evidence_quote":"Provides the Slice Wasserstein Distance implementation used to compute point cloud dissimilarities."},{"cited_title":"Seeds: Superpixels extracted via energy-driven sampling","cited_arxiv_id":null,"evidence_quote":"SEEDs algorithm used to generate the superpixels that define the annotation regions."}],"review_version":1}