{"id":"9373d4aa-b8fa-4c24-861d-33a56b4edf0b","arxiv_id":"2507.17367","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Adding a spatial diversity objective to region-based active learning improves semantic segmentation accuracy at low annotation budgets.","lead":"This paper introduces a spatial diversity term for region-based active learning in semantic segmentation, so that the system asks for labels from image regions spread widely across the image instead of clustering around uncertain spots. The authors show that combining the standard uncertainty, entropy, with this spatial diversity reaches 95 percent of fully supervised performance using only 5 to 9 percent of labeled pixels.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The proposed spatial-diversity term may act mainly as image-level coverage rather than local label diversity; the paper reports no direct check of the mechanism, so the stated contribution is not yet isolated.","rationale":"I read the paper in good faith. The controlled comparisons, with a shared initial batch and training protocol, are a real strength, and the ablations of τ, a/b/c, N, and Max-Min versus Max-Sum are thorough. The abstract's 'outperforming all state-of-the-art' claim is overbroad because only BALD and RAL are compared externally and RAL's number is imported from another paper with a different backbone; the reader already flagged this. I focused instead on the mechanism because the paper's contribution is explicitly 'spatial diversity as a new objective,' not merely a performance heuristic. The piecewise distance in Eq. (6) makes cross-image distances equal to far same-image distances, so the optimization may reduce to selecting from as-yet-uncovered images. This is a concrete, testable confound: a simple image-coverage baseline with entropy would isolate the effect. The direct correlation test would settle whether the proxy assumption in Section III-B is valid. Neither test is expensive relative to the experiments already reported. The empirical headline could survive either outcome, but the paper's framing and the claimed novelty should be adjusted if the mechanism check fails. Therefore I recommend a conditional acceptance rather than an unconditional one.","tokens_in":15587,"tokens_out":17019,"duration_ms":178982,"concrete_test":"Run an ablation on both datasets: replace the piecewise spatial term in Entropy+Spatial with a hard image-level constraint that selects, at each greedy step, the highest-entropy remaining region whose image is not yet represented in Lt∪Bt, using the same batch schedule and training protocol. Compare mIoU at all six budgets. Separately, on ground-truth maps, compute the Spearman correlation between ds(xi,xj) of Eq. (6) and a label-diversity measure, e.g., 1 minus the intersection-over-union of per-region class histograms, for same-image region pairs. If the image-coverage baseline matches or beats Entropy+Spatial, or if the correlation is near zero within either dataset, the central mechanism claimed in Section III-B is not supported and the paper should be reframed.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central empirical claim, that Entropy+Spatial improves over Entropy and feature-diversity baselines on Cityscapes and PASCAL VOC, is supported by the controlled experiments. The load-bearing uncertainty is about the mechanism. In Eq. (6) the piecewise spatial distance assigns distance c=2 to regions in different images and b=2 to same-image regions farther than τ; only same-image pairs within τ receive the smaller penalty a=1. Consequently, once any region of an image is labeled, the spatial term penalizes only nearby same-image candidates; regions in all other images receive the same maximum distance and are not differentiated by the spatial objective. The greedy selection therefore mostly maximizes the number of images covered, and the 'local spatial diversity' story in Section III-B is not the only explanation. Table IV shows Entropy+Spatial covers more images than Entropy or Entropy+Feature and approaches Random's coverage, but no baseline separates 'one high-entropy region per image' from the piecewise local penalty. The paper also does not report a direct correlation between ds and label diversity, leaving the proxy assumption in Section III-B untested. If image-level coverage is the active ingredient, the claimed novelty of a local spatial-distance function and the rationale based on category diversity are not established, even though the method may still work.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Entropy+Spatial, a region-based active learning method for semantic segmentation that augments an entropy uncertainty term with a coordinate-based spatial diversity term inside a max-min batch-selection objective solved greedily. The spatial distance is a piecewise constant function (Eq. 6) that penalizes nearby same-image regions and assigns equal distance to far same-image and cross-image regions. Experiments on Cityscapes and PASCAL VOC 2012 compare against Random, Entropy, Core-Set, Entropy+Feature, and combined variants, and include ablations of the distance form, tau, a, b, c, region size, objective form, feature dimension, runtime, and image coverage. The method reports about 95-96% of fully supervised mIoU with 8.4%/5.9% labeled pixels and a fixed-budget improvement over RAL.","tokens_in":15821,"tokens_out":12218,"duration_ms":125952,"significance":"If the empirical results hold, the method is practically valuable: it replaces feature-based diversity with a cheap coordinate-only penalty, is compatible with uncertainty sampling, and scales to about 10^5 regions. The paper has notable strengths: controlled comparisons with 3-run means and standard deviations, a broad ablation suite, a computational cost analysis, and concrete performance claims. The central relative improvement over Entropy and Entropy+Feature appears robust. However, the formal statement of the optimization objective is flawed, and the mechanism behind the spatial term is not fully isolated from image-level coverage, so the conceptual contribution needs additional support before the paper can be accepted as is.","major_comments":[{"comment":"The objective as displayed is not well posed. Written literally, min_{x_k in B_t, x_i,x_j in L_t union B_t} [lambda_u u(x_k)+d(x_i,x_j)] allows x_i=x_j, making d=0, and the pairwise term is independent of x_k; the minimum can collapse to the entropy-only term. The greedy criterion in Eq. (4) is coherent, but it is not the reformulation claimed in the text. Please restate Eq. (3) in the form max_{B_t} min_{x_k in B_t} [lambda_u u(x_k) + min_{x_i in L_t union B_t, x_i != x_k} d(x_i,x_k)] (or an equivalent max-min facility-dispersion objective) and adjust the set-function definitions in Eq. (2) accordingly.","section":"Section III-A, Eq. (3)"},{"comment":"With the default a=1, b=2, c=2, Eq. (6) assigns the same distance (2) to any candidate in an image without a selected region and to any same-image candidate farther than tau from the selected set; only same-image candidates within tau receive the smaller penalty. Consequently, the objective contains a strong image-coverage component, and the claimed 'local spatial diversity' mechanism is not isolated. The ablation in Section V-C showing that varying c/b has little effect argues against image coverage being the sole driver, but the paper never compares Entropy+Spatial against an image-coverage-only baseline or a local-penalty-only baseline, nor does it report a direct relationship between d_s and label diversity. Please add such ablations (for example, entropy plus a constant per-image penalty, and entropy plus a same-image-only local penalty) and/or measure the label diversity of the selected sets.","section":"Section III-B, Eq. (6), and Section IV-F"},{"comment":"The fixed-budget comparison with RAL is imported from [7] rather than rerun under the paper's training protocol, and RAL's budget includes 350 labeled images used to train its policy. The text's 'for fairness' claim is not fully supported: differences in training iterations, augmentation, and pre-training may affect the comparison. The abstract's claim of outperforming all state-of-the-art region-based AL methods rests partly on this table. Either rerun RAL under the same protocol or qualify the claim to the controlled experiments in Fig. 2.","section":"Section IV-D, Table I"}],"minor_comments":[{"comment":"The percentage targets are inconsistent: the abstract and conclusion say 95% with 8.4%/5.9% labeled pixels, while Section IV-D says 'more than 96%' at 8.4% on Cityscapes. Please align the claims.","section":"Abstract/Conclusion vs. Section IV-D"},{"comment":"Eq. (5) defines d_s using coordinates loc_i but does not state how coordinates are normalized or how distances between regions in different images are computed. Please clarify, since the linear variant in Table II is compared with the piecewise variant.","section":"Section III-B, Eq. (5)"},{"comment":"Algorithm 1 uses 'while t <= T' with T as the maximum number of batches. If batches are indexed 0 through 5, the loop should likely be t < T, or T should be defined accordingly.","section":"Algorithm 1"},{"comment":"Table IV reports the number of annotated images but does not explicitly state that the numbers are for Cityscapes; please state the dataset and, if available, provide the PASCAL VOC counterpart or explain its omission.","section":"Table IV"},{"comment":"There is a typo in the fully supervised training details: '0.007 for VOC 2012 and .' should be completed.","section":"Section IV-B"}],"recommendation":"major_revision","confidential_remarks":"This is a borderline case. The empirical contribution is solid, and the image-coverage concern is partly mitigated by the c/b ablation in Section V-C, which suggests the local penalty, not image coverage, is the active ingredient. However, the formal statement of Eq. (3) is genuinely incorrect as written, and the mechanism isolation requested in major comment 2 is important for the paper's conceptual claim. A major revision is appropriate; rejection would be too harsh given the controlled experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the empirical result is real. Adding the spatial term to entropy-based region selection improves region-based active learning for semantic segmentation on Cityscapes and PASCAL VOC, and the experiments are careful enough that I trust the numbers. The paper does not, however, isolate the mechanism it claims. In Eq. (6), the piecewise distance gives the same value (c=2) to any region in a different image as to a same-image region farther than tau. So once the first region in an image is selected, the spatial term stops differentiating among the remaining images. Greedy selection therefore mostly maximizes how many images get at least one region labeled; local in-image spread matters only for later picks within the same image. Table IV confirms that Entropy+Spatial covers more images than Entropy or Entropy+Feature. The paper never directly tests whether selected regions actually have higher label diversity, so the 'local spatial diversity' story in Section III-B is not established. This does not sink the method, but the contribution may be better described as encouraging image-level coverage with a local discouragement term.\n\nWhat is genuinely new: the first explicit spatial-diversity objective for region-based AL, the max-min reformulation that frames Core-Set and USDM as special cases, and a distance function cheap enough to make selection about 20x faster than feature-based diversity. The ablation coverage is thorough: distance function form, tau, a/b/c ratios, region size, max-min vs max-sum, feature dimension, and the paper reports three-run means with standard deviations. That is above average for this literature.\n\nSoft spots, in proportion. The abstract's claim to 'outperform all state-of-the-art region-based active learning methods' is too strong: the only external comparisons are RAL and BALD at one budget, and the RAL numbers are imported from the original paper. No code is released. Hyperparameters are tuned on the same datasets used for evaluation, which is mild but real. The biggest issue is the underdetermined mechanism, since an alternative 'one region per image' baseline would clarify whether the local distance term adds anything beyond coverage.\n\nThis paper is for people working on active learning for dense prediction, and it is a useful practical reference. I would send it to reviewers, asking for a direct test of the image-coverage hypothesis and code release. Not a desk reject.","headline":"The empirical result is solid, but the proposed spatial-diversity mechanism likely reduces to image-level coverage; still a useful, well-ablated contribution that deserves peer review.","tokens_in":16367,"tokens_out":3799,"would_cite":true,"duration_ms":36977,"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":"Spread-out regions reach 95% of full supervision with 9% of pixels","keywords":["active learning","semantic segmentation","region-based active learning","spatial diversity","entropy sampling","max-min optimization","label efficiency"],"falsifier":"A controlled experiment that permutes region coordinates while keeping the entropy values fixed should make the spatial term's benefit disappear if the benefit comes from the coordinate-label correlation; if Entropy+Spatial still beats Entropy after such permutation, the spatial term is doing something else. Concretely, run the paper's Cityscapes protocol with region locations randomly shuffled across images and compare mIoU curves.","tokens_in":15373,"feed_emoji":"🎯","tokens_out":5720,"duration_ms":55109,"temperature":0.7,"pith_summary":"This paper argues that when active learning labels image regions instead of whole images, selected regions should be spread out in image space as well as uncertain. Nearby regions in a segmentation image tend to share labels, so a batch chosen only by uncertainty is redundant. The authors add a spatial-diversity term, computed purely from region coordinates, to a max-min selection objective and solve it greedily. On Cityscapes and PASCAL VOC 2012, the resulting Entropy+Spatial method reaches about 95% of the accuracy of a fully supervised model while labeling only 8.4% and 5.9% of pixels, respectively.","feed_headline":"Spread-out regions reach 95% of full supervision with 9% of pixels","feed_subtitle":"Spatial diversity plus entropy matches full supervision using only 5 to 9 percent of labeled pixels.","key_machinery":"The load-bearing object is the piece-wise constant spatial distance $d_s(x_i,x_j)$, which assigns value $a$ to regions within distance $\\tau$ in the same image, $b$ to regions farther apart in the same image, and $c$ to regions from different images, with $c \\ge b \\ge a > 0$ and $b \\le 2a$ so it is a metric. This distance is inserted into a max-min objective $\\max_{B_t} \\min[\\lambda_u u(x_k) + d(x_i,x_j)]$ in which the unary term is per-region entropy and the pairwise term is diversity. A greedy algorithm selects, at each step, the region maximizing the minimum over all already selected regions of entropy plus distance, which scales to the roughly $10^5$ regions in benchmark segmentation datasets.","core_discovery":"The central claim is that spatial diversity is a useful objective for region-based active learning, and can replace feature-space diversity for semantic segmentation. The paper shows that combining per-region entropy with a piece-wise spatial distance, which penalizes selecting nearby regions while treating far-apart and cross-image regions equally, improves over entropy alone and over feature-diversity baselines such as core-set and entropy-plus-feature in the reported experiments. Because the spatial distance depends only on region coordinates, it remains reliable when the feature extractor is poorly trained or when feature distances concentrate in high dimensions. The authors also report that the method outperforms a reinforcement-learning region selector at a smaller labeling budget.","pith_inferences":["If spatial distance is a proxy for semantic diversity, the method should transfer to settings where categories correlate with location, such as street scenes, but lose its edge on datasets with objects uniformly repeated everywhere; that is an empirical prediction the paper does not test.","One could make the spatial penalty adaptive by shrinking $\\tau$ as the labeling budget grows, which the paper's own sensitivity experiments suggest may help late-stage selection.","The same max-min objective could be applied to whole-image or video-frame selection, where 'spatial' becomes temporal or geometric distance, a direct extension of the paper's argument.","Because the method needs no feature extractor and no extra labeled data, it is a natural plug-in for settings with domain shift, where a pretrained feature extractor would be unreliable."],"forward_implications":["Region-based active learning can reach near-fully-supervised accuracy with a few percent of labeled pixels, so annotation budgets for semantic segmentation can shrink by an order of magnitude.","Spatial diversity removes the need for a trained feature extractor during selection, making each active-learning round roughly 20 times faster than feature-diversity selection on Cityscapes in the paper's measurements.","The spatial term is compatible with uncertainty and feature diversity and can be added to existing selection objectives without extra training data.","Because the objective is a max-min diversification problem, the same greedy solver applies to any region representation, including superpixels or irregular patches.","The paper's fixed-budget comparison suggests that a simpler explicit selection rule can beat a learned reinforcement-learning policy while using less annotation budget."],"supporting_citations":[{"why":"Supplies the core-set feature-diversity baseline that the paper's spatial objective is designed to replace and the Eq. (1) instantiation.","marker":"[4]"},{"why":"Defines the uncertainty-plus-diversity formulation that the paper generalizes as USDM in Eq. (1).","marker":"[2]"},{"why":"Provides the result-diversification max-min formulation and greedy algorithm used for active selection in Eq. (3).","marker":"[24]"},{"why":"Provides the reinforced active learning baseline and its reported mIoU at 12k regions used in the fixed-budget comparison.","marker":"[7]"},{"why":"Establishes the region-based active learning setting for semantic segmentation that this work builds on.","marker":"[6]"},{"why":"BALD is a baseline in the fixed-budget comparison and represents Bayesian uncertainty active learning.","marker":"[3]"},{"why":"Prior region-based active learning with entropy on superpixels, showing spatial diversity was unexplored in that setting.","marker":"[21]"}],"fun_headline_variants":["Spatial diversity hits 95% of full supervision with 5-9% pixels","Spatial spread beats feature diversity: 95% with 5-9% pixels","Region-based active learning: spatial diversity yields 95% with 5-9% pixels","95% of full supervision from just 5-9% pixels via spatial diversity","Spatial diversity reaches 95% of full supervision with 5-9% pixels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central assumption is that image coordinates are a reliable stand-in for label diversity: regions that are far apart are assumed to contain different semantic categories, so spreading selections across the image removes redundancy.","fun_headline_variants_meta":{"raw":{"variants":["Spatial diversity hits 95% of full supervision with 5-9% pixels","Spatial spread beats feature diversity: 95% with 5-9% pixels","Region-based active learning: spatial diversity yields 95% with 5-9% pixels","95% of full supervision from just 5-9% pixels via spatial diversity","Spatial diversity reaches 95% of full supervision with 5-9% pixels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00168,"raw_usage":{"total_tokens":6611,"prompt_tokens":849,"completion_tokens":5762,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":465,"completion_tokens_details":{"reasoning_tokens":5648}},"tokens_in":465,"tokens_out":5762,"duration_ms":40332,"temperature":1.0,"reasoning_tokens":5648,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:48:33.194016+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled experiment that permutes region coordinates while keeping the entropy values fixed should make the spatial term's benefit disappear if the benefit comes from the coordinate-label correlation; if Entropy+Spatial still beats Entropy after such permutation, the spatial term is doing something else. Concretely, run the paper's Cityscapes protocol with region locations randomly shuffled across images and compare mIoU curves.","supporting_citations":[{"cited_title":"Active learning for convolutional neural networks: A core-set approach,","cited_arxiv_id":null,"evidence_quote":"Supplies the core-set feature-diversity baseline that the paper's spatial objective is designed to replace and the Eq. (1) instantiation."},{"cited_title":"Multi-class active learning by uncertainty sampling with diversity maximization,","cited_arxiv_id":null,"evidence_quote":"Defines the uncertainty-plus-diversity formulation that the paper generalizes as USDM in Eq. (1)."},{"cited_title":"An axiomatic approach for result diversi- fication,","cited_arxiv_id":null,"evidence_quote":"Provides the result-diversification max-min formulation and greedy algorithm used for active selection in Eq. (3)."},{"cited_title":"Rein- forced Active Learning for Image Segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the reinforced active learning baseline and its reported mIoU at 12k regions used in the fixed-budget comparison."},{"cited_title":"CEREALS - cost-effective region-based active learning for semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"Establishes the region-based active learning setting for semantic segmentation that this work builds on."},{"cited_title":"Deep bayesian active learning with image data,","cited_arxiv_id":null,"evidence_quote":"BALD is a baseline in the fixed-budget comparison and represents Bayesian uncertainty active learning."},{"cited_title":"Region-based active learning for efficient labeling in se- mantic segmentation,","cited_arxiv_id":null,"evidence_quote":"Prior region-based active learning with entropy on superpixels, showing spatial diversity was unexplored in that setting."}],"review_version":1}