{"id":"9024170d-b5e5-4b4b-acfb-f821bf1d339e","arxiv_id":"2501.13529","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A few-shot segmentation framework that estimates each support image's contribution and prunes low-contribution supports, combined with symmetric correlation, maintains or improves mIoU as support count grows to 70 on COCO-20i and PASCAL-5i.","lead":"This paper identifies 'support dilution' in few-shot semantic segmentation: adding more support images can hurt accuracy, and proposes a method with a contribution index, symmetric correlation, and support pruning to keep only useful supports. The result matters because real-world systems often collect large, noisy support sets automatically, and this work aims to make segmentation robust without manual filtering.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 8's Jensen lower bound drops the softmax partition function and is unverified; pruning may not select true high-contributed supports.","rationale":"I read the paper as claiming that support dilution is a real failure mode and that the three components—δ, SC, and pruning—fix it. The empirical evidence, including plug-and-play gains and cross-domain improvements over DCAMA/MSANet, is meaningful even without released code. But the mechanism for pruning is theoretically insecure: Eq. 8 is presented as a Jensen lower bound, yet the softmax partition function and the max operation disappear without stated assumptions, and the resulting θ in Eq. 9 is what the greedy algorithm actually optimizes. This is an internal derivation gap, not merely a disagreement with prior work. The most direct way to settle it is oracle-based selection quality: if the pruned subset agrees with leave-one-out mIoU-defined importance, the method survives regardless of Eq. 8's status; if not, the central 'select high-contributed supports' claim fails. I also note a minor empirical inconsistency: Table VI reports a small N=50 dip for +SC (54.2 vs 55.4 at N=30), so the verbatim claim that mIoU 'keeps climbing from 1 to 70 shots' should be softened unless Fig. 7's fold-averaged curves hide that dip. None of this requires questioning authorial integrity; it is an argument-support gap. Conditional acceptance remains the right disposition: release code, verify pruning selection against an oracle, and repair or qualify Eq. 8.","tokens_in":20063,"tokens_out":7039,"duration_ms":70958,"concrete_test":"On COCO-20i fold 1, take the 70-shot support pools used in Fig. 7 and define an oracle contribution for each support by leave-one-out inference: run the full pipeline with that support removed and record the query mIoU drop. Compare the oracle top-30 supports with the N'=30 subset selected by Support Image Pruning, and also compare the actual sum of δ over the pruned subset versus the oracle subset. If the overlap is near chance or the pruned subset's real δ sum is below the oracle subset's, the pruning mechanism is not selecting what it claims, and the Eq. 8 derivation cannot be rescued by appeal to empirical success.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-E's pruning stage is the component that turns a large noisy pool into a compact high-quality set, so its selection criterion must rank supports by true contribution. Eq. 8 is the hinge, and it is not derived: the LHS is the mean over support tokens of a max softmax attention weight, while the RHS is a raw dot product between mean normalized features, with the softmax partition function and the max removed. Jensen's inequality alone does not imply this bound; it would need additional assumptions about f1 magnitudes and partition-function normalization, which the paper neither states nor checks. Eq. 9 then replaces the stated objective (maximize sum of δ) with this lower bound, so Algo. 1 optimizes θ rather than the contribution index it claims to maximize. The large-N gains attributed to pruning therefore rest on an unverified surrogate. Neither Eq. 4's δ nor the Eq. 9 bound is validated against any ground-truth notion of support contribution; Table VII only compares θ against generic L2/cosine retrieval and does not include an oracle or random baseline, so it cannot show whether true high-contributed supports are being selected.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper addresses the problem of \"support dilution\" in few-shot semantic segmentation (FSS), where increasing the number of support images can degrade rather than improve segmentation quality. The authors propose three components: a contribution index δ (Eq. 4) based on maximum softmax attention weights; a Symmetric Correlation (SC) module that uses a shared normalization f(x)=f1(x)f2(x)/||f2(x)|| for keys and queries (Eqs. 5-6); and Support Image Pruning, a greedy subset-selection algorithm based on an approximate objective θ (Eq. 9) derived from δ. Experiments on COCO-20i and PASCAL-5i with ResNet-50, ResNet-101, and Swin-B backbones show consistent mIoU improvements over DCAMA and MSANet as N grows from 1 to 70, together with cross-domain, plug-and-play, online, and real-world demonstrations.","tokens_in":20313,"tokens_out":6174,"duration_ms":58389,"significance":"The support-dilution problem is practically important and has received little attention, and the paper gives a clear motivation with extensive experiments. If the pruning derivation can be made correct or empirically validated, the method would be a valuable contribution: the plug-in gains over HDMNet, MSANet, and DCAMA and the reduced inference cost at large N are concrete and useful. The manuscript is generally well written and transparent about its training setup, and the online and real-world experiments strengthen the practical story. However, the theoretical justification of the pruning objective is currently not sound, and the contribution-index validation is incomplete, so the central claims cannot be fully accepted without revision.","major_comments":[{"comment":"Equation (8) is not a valid Jensen bound as stated. The left-hand side is an average over support tokens of a maximum softmax attention probability, while the right-hand side is a raw dot product of f applied to averaged tokens. The softmax partition function and the max operation are removed, and no assumptions on the range or distribution of the logits are given that would make the inequality hold. Because Eq. (9) and Algorithm 1 replace the stated objective in Eq. (7) with this lower bound, the pruning stage currently optimizes a different and unvalidated surrogate rather than the contribution index it claims to maximize. Moreover, δ in Eq. (4) depends on the whole support set through the softmax denominator, so the objective in Eq. (7) is not a simple sum of per-support independent terms; the greedy algorithm does not solve Eq. (7). Please provide a correct derivation with explicit assumptions, or recharacterize Eq. (9) as an empirical heuristic and validate it directly against δ and against segmentation performance, for example by comparing selected subsets with those chosen by an oracle.","section":"IV-E, Eq. (8)"},{"comment":"The claim that using the same network f for keys and queries \"guarantees the attention function to have an unique maximum point\" for identical support-query pairs is not proved and does not follow from symmetry alone. A symmetric score f(x)^T f(y) can have f(x)^T f(z) >= f(x)^T f(x) for some z different from x unless f is sufficiently injective and the magnitude term f1 is positive; the normalization f(x)=f1(x)f2(x)/||f2(x)|| does not by itself ensure this. Furthermore, the upper-bound experiments in Fig. 4 and Table I partly test a property that SC is explicitly designed to enforce, so those results should be interpreted as a sanity check of the architecture rather than as independent evidence that high-contributed supports are being recognized. The benchmark improvements in Fig. 7 remain the primary evidence for the method.","section":"IV-D"},{"comment":"The contribution index δ is defined purely from attention weights and is never checked against any ground-truth notion of support contribution. Table VII compares Support Image Pruning with feature-distance retrieval baselines but does not include an oracle baseline (for example, selecting supports by measured segmentation gain or by the known upper-bound support) or a random-selection baseline, so the experiments cannot show whether the pruned subset actually contains the true high-contributed supports. Adding such baselines, or an ablation in which the contribution ranking is corrupted or permuted, would directly test the validity of δ and of the pruning objective.","section":"IV-C, Table VII"},{"comment":"All benchmark numbers are reported as single point estimates without error bars or multiple seeds. Some of the claimed gains at small N (for example, 1-shot in several folds of Fig. 7) are small and could be within run-to-run variation; given the headline claim that mIoU \"keeps climbing from 1 to 70 shots\", please report means and standard deviations over at least three random seeds for the main comparisons, or state the number of runs used and provide a variance analysis.","section":"V-C, Fig. 7"}],"minor_comments":[{"comment":"The word \"backpropogation\" should be \"backpropagation\".","section":"V-B"},{"comment":"The caption contains the typo \"Transfromer\"; it should be \"Transformer\".","section":"Fig. 8 caption"},{"comment":"The section text uses \"Plug-and-paly\"; this should be \"Plug-and-play\".","section":"V-E"},{"comment":"Reference [12] is cited in the text as the source of ResNet, but the bibliography entry is for Hariharan et al.; please update the reference to the original ResNet paper.","section":"References"},{"comment":"The softmax expressions contain an extra closing parenthesis after the argument; please clean up the notation for clarity.","section":"Eqs. (3), (6), (12)"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely problem and reports large, internally consistent gains, which makes it attractive for the FSS community. My main concern is the unproven and likely incorrect Jensen step in Eq. (8), which is load-bearing for the pruning component. If the authors can either prove the inequality under stated assumptions or reframe pruning as an empirical heuristic with direct validation against δ and oracle selection, the paper could become acceptable. I would also encourage the editor to ask for error bars and code release, and to have the authors address the circularity concern in the upper-bound diagnostic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the paper. The support-dilution phenomenon is real and worth studying, and the paper does a good job documenting it: DCAMA and friends lose mIoU as N grows past roughly 5–10, and the upper-bound support experiment makes the point visually. That systematic study is the main new contribution. The plug-in experiments (Tab. IV) are also genuinely informative: adding their SC and pruning to DCAMA, HDMNet, and MSANet gives consistent large-N gains, which suggests the components transfer. Credit where due: this is a practical problem with a plausible engineering fix.\n\nThe soft spots are concentrated in the pruning theory and validation. Eq. 8 claims Jensen's inequality, but the LHS is a mean of max softmax attention weights and the RHS is a raw inner product of normalized mean features; neither the softmax partition function nor the max survives. Jensen alone does not give that bound, and the paper never states the extra assumptions. So Eq. 9 optimizes a surrogate, not the stated contribution sum, and Algo. 1 inherits that. This does not necessarily kill the method — the surrogate might still work — but it means the paper's justification for the central large-N mechanism is unverified. The stress-test note is right about this.\n\nAlso, δ is never validated against any ground-truth notion of contribution, and Tab. VII lacks an oracle or random baseline, so we cannot tell whether pruning actually selects high-contributed supports or just does feature retrieval. The upper-bound experiment is partly circular: SC is designed so an identical support-query pair gets maximal attention, so the upper-bound support resisting dilution is partly a property built into the architecture. The benchmark gains are still independent evidence, so I would not call the paper fatally flawed, but the evidence is weaker than the claims.\n\nMinor points: no error bars, code promised after acceptance, the main N-shot comparison in Fig. 7 covers only DCAMA and MSANet even though HDMNet and SCCAN appear in Tab. I, and the baselines are not retrained under a unified protocol. Training for 5 epochs off a DCAMA init is fine, but it makes SOTA claims fragile.\n\nBottom line: this deserves a serious referee, but I would send it back for major revision. The phenomenon and plug-in results are valuable; the pruning derivation needs to be fixed or honestly relabeled as a heuristic, and the contribution-index validation needs a random/oracle comparison.","headline":"Solid study of a real phenomenon, with a load-bearing derivation gap in the pruning step; worth reviewing, not worth accepting as-is.","tokens_in":20770,"tokens_out":2442,"would_cite":true,"duration_ms":23572,"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":"Support dilution in few-shot segmentation can be diagnosed with a contribution index and fixed with symmetric correlation and pruning.","keywords":["few-shot semantic segmentation","support dilution","contribution index","symmetric correlation","support image pruning","cross-attention","multi-shot segmentation"],"falsifier":"Compute an oracle contribution for each support by removing it, re-running segmentation, and measuring the mIoU change on a fixed set of queries. If the paper's contribution index has low rank correlation with that oracle, or if mIoU stops climbing when support pools are deliberately contaminated with many irrelevant images, the central claim is refuted.","tokens_in":19884,"feed_emoji":"📈","tokens_out":5930,"duration_ms":51737,"temperature":0.7,"pith_summary":"This paper tries to establish that a common failure mode in few-shot semantic segmentation, called support dilution, can be diagnosed and fixed. Support dilution is what happens when adding more example (support) images stops improving, and then starts hurting, segmentation of the query image: the useful supports get drowned out by irrelevant ones. The proposed remedy has three parts: a contribution index that scores each support by its attention to the query, a Symmetric Correlation module that makes the best-matching support dominate the attention, and a pruning operation that discards low-scoring supports before correlation is computed. If the paper is right, an FSS system can take a large, noisy support pool and keep getting better masks as the pool grows, which is exactly what real-world use on web-collected images would require.","feed_headline":"More supports, better masks: 70-shot beats 5-shot","feed_subtitle":"A contribution index, symmetric correlation, and pruning keep noisy examples from drowning useful ones.","key_machinery":"The load-bearing object is the contribution index $\\delta(x_s)$, defined as the average over a support's tokens of their maximum attention weight to any query token; it is meant to quantify how much a support truly helps a query. Symmetric Correlation rewrites the attention score so that keys and queries share one function $f(x)=f_1(x)\\frac{f_2(x)}{\\|f_2(x)\\|_2}$, with a magnitude part for objectness and a normalized angle part for similarity; symmetry makes the score maximal if and only if support and query are identical. Support Image Pruning turns subset selection into a greedy maximization of a Jensen-derived lower bound on the summed contribution index, reducing cost from enumerating all subsets to $O(N'\\times N)$. These three pieces work together: the index measures dilution, the symmetric attention resists it, and pruning removes the worst diluters before attention is applied.","core_discovery":"The central claim is that support dilution is a real and measurable effect, and that it can be overcome by coupling support scoring with a symmetric attention design. The paper shows that in existing methods the deviation between the contribution index of a perfect \"upper-bound\" support and the average of the other supports shrinks as the number of supports grows, and that segmentation mIoU falls with it. Its Symmetric Correlation uses the same network to produce keys and queries plus a normalization into magnitude and angle parts, so that attention is maximized exactly when support and query features match; this keeps high-contributed supports standing out even in a noisy pool. Support Image Pruning then greedily keeps the supports that maximize a lower bound on total contribution, which the paper justifies via Jensen's inequality, and the full pipeline reports mIoU that keeps climbing from 1 to 70 shots on both COCO-20i and PASCAL-5i across three backbones.","pith_inferences":["The paper's diagnosis suggests a general stress test for any attention-based few-shot learner: record how the contribution deviation changes as the support pool grows; a falling deviation predicts vulnerability to dilution even before mIoU drops.","Because the contribution index is defined purely from attention weights, an obvious extension is to validate it against an oracle contribution measured by leave-one-out changes in mask quality; the pruning objective could then be replaced or calibrated by that oracle.","The same dilution mechanism likely applies to other tasks where many weakly relevant examples are pooled, such as few-shot object detection or retrieval-augmented generation, wherever soft attention aggregates a large noisy context.","The Jensen lower bound used for pruning drops the softmax partition function, so the pruning score is unnormalized; a testable variant would divide by the partition or normalize by feature magnitude to see whether pruning choices change."],"forward_implications":["Segmentation quality continues to rise as the support count grows from 1 to 70 shots, instead of flattening or dropping.","The Symmetric Correlation and pruning components can be inserted into prior FSS methods and improve their multi-shot results, suggesting the fix is architectural rather than benchmark-specific.","When the support pool exceeds 30 images, pruning to 30 supports keeps the gains while holding inference time and memory roughly constant.","The method transfers across domain gaps and tolerates automatically generated support masks, so web-searched supports can be used without manual filtering."],"supporting_citations":[{"why":"Main SOTA baseline that exhibits support dilution and supplies the checkpoint used to initialize the pipeline.","marker":"[37]"},{"why":"Comparison baseline in the N-shot plots and a target for plug-and-play evaluation.","marker":"[16]"},{"why":"Baseline in the preliminary dilution study and a plug-and-play target.","marker":"[26]"},{"why":"Additional baseline in Table I showing mIoU decline as shots grow.","marker":"[51]"},{"why":"COCO-20i benchmark supplying the primary N-shot evaluation folds.","marker":"[27]"},{"why":"PASCAL-5i benchmark supplying the second evaluation and cross-domain test set.","marker":"[36]"},{"why":"Swin Transformer backbone used in experiments.","marker":"[24]"}],"fun_headline_variants":["More support images can hurt: new method fixes that","Stop noisy supports from ruining few-shot segmentation","Support dilution solved: selective support keeps mIoU climbing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the contribution index, which is computed from attention weights alone, correctly ranks how much each support truly helps segment the query; if that ranking is wrong, pruning will discard the useful supports and Symmetric Correlation will amplify the wrong ones.","fun_headline_variants_meta":{"raw":{"variants":["More support images can hurt: new method fixes that","Stop noisy supports from ruining few-shot segmentation","Support dilution solved: selective support keeps mIoU climbing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000728,"raw_usage":{"total_tokens":3278,"prompt_tokens":980,"completion_tokens":2298,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":2248}},"tokens_in":596,"tokens_out":2298,"duration_ms":16934,"temperature":1.0,"reasoning_tokens":2248,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:52:02.193834+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute an oracle contribution for each support by removing it, re-running segmentation, and measuring the mIoU change on a fixed set of queries. If the paper's contribution index has low rank correlation with that oracle, or if mIoU stops climbing when support pools are deliberately contaminated with many irrelevant images, the central claim is refuted.","supporting_citations":[{"cited_title":"In: European Conf","cited_arxiv_id":null,"evidence_quote":"Main SOTA baseline that exhibits support dilution and supplies the checkpoint used to initialize the pipeline."},{"cited_title":"In: IEEE/CVF Int","cited_arxiv_id":null,"evidence_quote":"Baseline in the preliminary dilution study and a plug-and-play target."},{"cited_title":"In: IEEE/CVF Int","cited_arxiv_id":null,"evidence_quote":"Additional baseline in Table I showing mIoU decline as shots grow."},{"cited_title":"In: IEEE/CVF Int","cited_arxiv_id":null,"evidence_quote":"COCO-20i benchmark supplying the primary N-shot evaluation folds."},{"cited_title":"In: IEEE/CVF Int","cited_arxiv_id":null,"evidence_quote":"Swin Transformer backbone used in experiments."}],"review_version":1}