{"id":"d0f9863b-5f91-4df7-931c-f006b159e848","arxiv_id":"2512.16415","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A training-free, three-stage exemplar-selection pipeline improves zero-shot object counting across natural, aerial, and medical images when compared with other inference-only methods.","lead":"CountZES is a no-training pipeline that picks three diverse example images (exemplars) for zero-shot object counting: one from detection, one from density, and one from feature consistency, then feeds them to a pretrained counter. It reports better cross-domain counting accuracy than several training-free methods, though it is not always best on the main benchmark as claimed.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CountZES's central 'superior performance' claim is contradicted by its own Table 1 on FSC-147.","rationale":"The reader correctly identified several weaknesses, including the FSC-147 MAE contradiction in the rationale, but selected SAM partial-mask failures as the weakest assumption. I argue that the mismatch between the paper's headline claim of 'superior performance' and its own quantitative results is the most load-bearing concern because it is a direct internal inconsistency: it does not depend on unresolved external limitations and, if not addressed, falsifies the central contribution as stated. The SAM partial-mask issue and the DGE self-referential loop are real and contribute to the conditional verdict, but the claim-evidence mismatch is more fundamental for the paper's significance. The verdict should remain CONDITIONAL, pending a rewording or statistical justification of the superiority claim.","tokens_in":17281,"tokens_out":11007,"duration_ms":100660,"concrete_test":"Directly compare the Abstract/Conclusion claim to Table 1. Tally the number of methods in Table 1 with lower MAE and RMSE than CountZES in both the in-domain and cross-domain settings. If any method (here, T2ICount, GeCo, CountGD; and CountingDINO in cross-domain) has lower values, the claim of 'superior performance' must be explicitly narrowed to 'competitive among inference-only zero-shot methods' or supplemented with statistical tests (e.g., paired bootstrap on FSC-147 test set) showing the differences are not significant. This settles whether the headline claim is supported by the paper's own data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and conclusion (Sec. 1, Sec. 5) claim 'superior performance compared to SOTA' and 'superior performance among ZOC methods.' Table 1 shows in-domain CountZES MAE 15.77, while T2ICount, GeCo, and CountGD achieve 11.76, 13.30, 14.76. In the cross-domain inference-only row, CountZES MAE 21.09 exceeds CountingDINO's 20.93 (and its RMSE 110.14 vs 71.37 is far higher), so even among inference-only methods the claim of lowest MAE/RMSE is false unless CountingDINO is excluded because it uses box prompts. The paper's own evaluation text uses 'competitive MAE' for in-domain results, which contradicts the strong wording of the central claim. This is an internal inconsistency, not a matter of external consensus, and it directly undermines the paper's stated contribution.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"CountZES proposes an inference-only pipeline for zero-shot object counting (ZOC) that discovers three exemplars per image through Detection-Anchored (DAE), Density-Guided (DGE), and Feature-Consensus (FCE) stages, then feeds them to a frozen pretrained density estimator. The method uses CLIP, GroundingDINO, SAM, and either CounTR or DSALVANet as fixed backbones, and claims superior ZOC performance while generalizing to natural, aerial, and medical domains. The paper reports results on FSC-147, CARPK, PerSense-D, VGG, and MBM, with ablations attributing gains to each stage.","tokens_in":17566,"tokens_out":4726,"duration_ms":46695,"significance":"If the headline claims held, the contribution would be a modular, training-free exemplar-selection strategy for ZOC, useful in settings where per-class fine-tuning is impractical. The paper is clearly written and includes sensible ablations, cross-domain protocols, and a runtime analysis. It also deserves credit for being explicit about its modularity and for reporting a limitations appendix. However, the central 'superior performance' claim is not supported by the paper's own main table on FSC-147, and the documented dependence on SAM's ability to produce complete single-instance masks imposes a stronger constraint than the abstract conveys. The significance of the work is moderate and conditional on a more careful framing of what is actually demonstrated.","major_comments":[{"comment":"The claim of 'superior performance among ZOC methods' is contradicted by Table 1 on FSC-147. In the in-domain setting, CountZES reports MAE 15.77, while T2ICount, GeCo, and CountGD report 11.76, 13.30, and 14.76 respectively. The paper's own Section 4 text describes the in-domain MAE as 'competitive,' which is accurate, but the abstract and conclusion use 'superior' without qualification. Even in the cross-domain row, CountingDINO* has lower MAE (20.93 vs 21.09) and much lower RMSE (71.37 vs 110.14). If CountingDINO is excluded because it uses box prompts, that exclusion must be stated wherever 'superior' appears. As written, the paper is internally inconsistent about its main claim.","section":"Abstract, Sec. 1, Sec. 5, and Table 1"},{"comment":"The DGE stage selects an exemplar using the same density estimator that later produces the final count. Density peaks (Eq. 4), the RoI-count single-instance filter (Eq. 5), the per-box count estimates (Eq. 10), and the pseudo-GT KDE mode (Eqs. 11-12) all come from the same frozen DE. Consequently, the DGE gain in Table 5 may reflect that an exemplar is consistent with the counter's own prior rather than that it is a better exemplar in any transferable sense. This is a circularity risk for the claim that DGE improves exemplar quality. Please provide an analysis where the selection-time DE differs from the final-count DE, or where exemplars selected by DGE are evaluated with an independent counter, to separate self-consistency from true representativeness.","section":"Sec. 3.2, Eqs. (4), (5), (10)-(12)"},{"comment":"The paper explicitly documents that SAM point prompts produce partial masks for watches, nail polishes, comic books, eggs, and sauce bottles, and states that 'residual noise can still propagate when the majority of candidate exemplars are imperfect.' This is a direct limitation on the central generalization claim, because all three stages rely on SAM masks as candidate boxes. Since several of these classes (e.g., eggs, comic books) appear in the benchmark classes used for evaluation, the unconditional wording of the abstract and conclusion is too strong. Please quantify how often partial-mask failures occur across the test splits and report per-class or per-infection performance, or restrict the claims to classes where SAM provides complete instances. A limitation appendix is welcome, but it must be reflected in the paper's central claims.","section":"Appendix A.7 and Figs. 13-14"},{"comment":"The cross-domain comparisons on PerSense-D and VGG/MBM include very few inference-only baselines: only TFOC on PerSense-D, and TFOC on VGG/MBM. Since the paper positions CountZES against inference-only methods, the absence of CountingDINO, TFCAC, OmniCount, and Count Anything from these tables makes it difficult to substantiate 'superior performance among inference-only methods' outside FSC-147. Please add the missing baselines or explicitly justify their exclusion.","section":"Sec. 4, Tables 3 and 4"}],"minor_comments":[{"comment":"The column header 'Density-based Class-wise' is confusing: the table appears to show density-based splits, not class-wise results. Also, the 'Overall' MAE for GeCo (13.12) is higher than each of its Low/Med/High MAE entries, which suggests an inconsistency in how the overall row is computed or labeled.","section":"Table 3"},{"comment":"Eq. (10) states \\hat{c}_i \\in \\mathbb{Z}, but the density estimator returns a continuous count (and Eq. (5) uses fractional predictions like 1<\\hat{c}<2). The codomain should be nonnegative reals.","section":"Appendix A.3, Eq. (10)"},{"comment":"The bibliography contains Langley (2000), but I do not see a citation to it in the text. Please either cite it or remove it.","section":"References"},{"comment":"Minor formatting issues: 'Figure.' appears with inconsistent spacing, 'V A-Count' and 'DA VE' formatting is awkward, and Table 1 has two 'CountZES (Ours)' rows that could be distinguished more clearly with explicit 'in-domain' and 'cross-domain' labels in the rows.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper's own Table 1 and Appendix A.7 undermine the unconditional 'superior performance' claim. The engineering contribution is real and the ablations are plausible, but the authors need to realign the claims with the evidence, address the circularity concern with an independent-counter experiment, and provide quantitative analysis of the SAM partial-mask failures. These are fixable within the scope of a revision, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: CountZES is a real, useful entry in zero-shot object counting — a three-stage, inference-only exemplar-selection pipeline with strong cross-domain results — but the paper sells itself as better than its own Table 1, and the density-guided stage has a self-referential loop that deserves scrutiny.\n\nWhat's actually new is the combination, not the parts. CLIP, GroundingDINO, SAM, density-peak sampling, and feature clustering are known; arranging them as Detection-Anchored, Density-Guided, and Feature-Consensus stages, each contributing one exemplar to a fixed density estimator, is a sensible and apparently effective scheme. The ablations show incremental gains (25.13 → 20.75 → 15.77 MAE), and the cross-domain numbers are the strongest part: on CARPK, CountZES gets 7.24 MAE against T2ICount's 8.61 and GeCo's 10.34; on the medical sets it beats both by large margins. The limitation section (A.7) honestly documents SAM fragmenting objects into partial masks, with figures. Credit where it's due.\n\nSoft spots, roughly in order.\n\nFirst, claim calibration. The abstract says \"superior performance among ZOC methods\" and the conclusion repeats it, but Table 1 shows in-domain FSC-147 MAE 15.77 against 11.76 for T2ICount and 13.30 for GeCo. The body text is more careful — \"competitive MAE and lowest RMSE\" is what the table actually supports. The stress-test charge about the cross-domain row (CountingDINO's 20.93 vs 21.09) mostly dissolves if you accept the asterisk exclusion: CountingDINO uses box prompts, so it isn't playing the same zero-shot game, and among genuinely prompt-only inference methods CountZES leads that row. But the exclusion needs to be argued, not just footnoted, and the abstract should not claim superiority over all ZOC methods when the standard benchmark says otherwise.\n\nSecond, DGE is partly self-referential. The same pretrained density estimator finds the density peaks, computes the RoI counts, produces the pseudo-GT as the KDE mode of those same counts, scores boxes by closeness to that pseudo-GT, and then receives the winning exemplar to produce the final count. Choosing an exemplar that agrees with the estimator's own prior is a reasonable heuristic, and the ablation shows it helps, but the paper should say plainly that the pseudo-GT is not an independent signal. As written, \"count-driven consistency\" reads stronger than the evidence supports.\n\nThird, the SAM partial-mask problem is real but documented, and it is confined to the appendix. Given that A.7 is one of the better sections, the main text should point to it and note which classes fail. Fourth, minor: the abstract promises code with no link, and most hyperparameters (α, w_sim/w_ent, the RoI interval) get no sensitivity analysis beyond the k=16 ablation.\n\nWho this is for: people working on class-agnostic or zero-shot counting. The engineering is solid, the cross-domain results are the headline, and the flaws are addressable. I would send it to review, with the authors asked to reconcile the abstract with Table 1 and add an honest paragraph on the DGE loop. A serious referee can get this into good shape.","headline":"Genuinely useful inference-only exemplar-selection pipeline for zero-shot counting with strong cross-domain results, but the abstract overstates 'superior performance' against its own Table 1 and the density-guided stage is self-referential.","tokens_in":18055,"tokens_out":8442,"would_cite":true,"duration_ms":74971,"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":"CountZES claims that zero-shot object counting can be done without any task-specific training by selecting three complementary exemplars—semantic, density-guided, and feature-consensus—from fixed pretrained models.","keywords":["zero-shot object counting","exemplar selection","inference-only","open-vocabulary detection","density estimation","point prompting","cross-domain generalization","foundation models"],"falsifier":"Take classes where the segmenter is known to split objects (e.g., watches segment into dial and band, eggs into yolk and white). If replacing the segmenter with one that yields complete masks does not reduce CountZES's error on those classes, then exemplar selection itself is not what is driving the reported accuracy; the same could be tested by measuring whether selected exemplar masks cover at least 80% of the annotated instance.","tokens_in":17143,"feed_emoji":"🔢","tokens_out":4912,"duration_ms":51178,"temperature":0.7,"pith_summary":"The paper tries to establish that counting instances of a category never seen during training, given only its name, can be solved by choosing good exemplars rather than by training a counter. It introduces CountZES, an inference-only pipeline that selects one exemplar from each of three complementary stages: text-conditioned detection, density-map consistency, and feature-space consensus. These three exemplars are fed together to a frozen pretrained density estimator, and the authors report that this selection strategy outperforms other inference-only methods across natural, aerial, and medical images. The central claim is that exemplar quality—not regression capacity—is the bottleneck in zero-shot counting, and that a deliberately diverse set of three exemplars suffices. If true, counting a new class would require no annotations or fine-tuning, only a class name and a few frozen foundation models.","feed_headline":"No retraining needed: three exemplars per image do the counting","feed_subtitle":"CountZES picks one exemplar from each of three stages, then counts with frozen pretrained models.","key_machinery":"The central object is the multi-stage exemplar selection pipeline, CountZES. The Detection-Anchored Exemplar (DAE) stage refines a detector box into a single-instance region by scoring SAM masks with text-alignment and entropy. The Density-Guided Exemplar (DGE) stage turns density-map peaks into point prompts for SAM, filters boxes by predicted region count, and selects the box whose count is closest to the mode of all candidate counts while remaining semantically compact. The Feature-Consensus Exemplar (FCE) stage clusters SAM feature embeddings of single-instance boxes and picks the one nearest the majority-cluster centroid. The mechanism is the deliberate diversity of the resulting three-","core_discovery":"On its own terms, the paper discovers that zero-shot object counting can be recast as an exemplar-inference problem governed by semantic, statistical, and feature-level priors. CountZES refines a coarse open-vocabulary detection into a single-instance exemplar using text-conditioned similarity peaks and a segmenter; it then derives a second exemplar from density-map peaks whose region count and semantic compactness agree with a pseudo-ground-truth mode; and it selects a third exemplar by clustering segmenter features and choosing the box nearest the majority-cluster centroid. The authors claim this one-exemplar-per-stage diversity is what drives accuracy, and they support the claim with expe","pith_inferences":["A direct corollary the paper leaves implicit is that per-class accuracy should be predictable from the segmenter's mask completeness for that class; this can be tested by correlating per-class MAE with manual mask-completeness scores.","The one-exemplar-per-stage rule is a fixed design choice; an adaptive weighting of stages based on detector confidence, density spread, or feature-cluster tightness could improve counts on images where one stage is unreliable.","The same exemplar-selection logic could extend to other dense prediction tasks, such as promptable instance segmentation or referring localization, where a text prompt must be converted into representative visual support without labels."],"forward_implications":["A user can count arbitrary categories—cars in aerial views, cells in microscopy—from a class name alone, with no exemplar annotation and no task-specific optimization.","Any improvement to a single component (detector, segmenter, density estimator) can be dropped into the pipeline directly, since all components remain fixed during inference.","Because text is used only as an inference-time probe, the method avoids the text-count entanglement that the paper identifies as a source of cross-domain brittleness in other zero-shot counters.","On the reported benchmarks, CountZES matches or beats several trained zero-shot counting methods while using only frozen pretrained components."],"fun_headline_variants":["Zero-shot counting: three exemplars, no retraining","Three-stage exemplar selection for zero-shot counting","CountZES: Better exemplars, zero-shot counting","Infer exemplars from text, density, and features","No training, just three exemplars: CountZES"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The entire pipeline presumes that a single point prompt reliably makes the segmenter return the whole object, not a fragment of it; when it returns only a dial, a yolk, or a character on a book cover, the selected exemplar is partial and the final count is conditioned on that partial object.","fun_headline_variants_meta":{"raw":{"variants":["Zero-shot counting: three exemplars, no retraining","Three-stage exemplar selection for zero-shot counting","CountZES: Better exemplars, zero-shot counting","Infer exemplars from text, density, and features","No training, just three exemplars: CountZES"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000262,"raw_usage":{"total_tokens":1448,"prompt_tokens":777,"completion_tokens":671,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":592}},"tokens_in":521,"tokens_out":671,"duration_ms":6908,"temperature":1.0,"reasoning_tokens":592,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T15:31:42.525140+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take classes where the segmenter is known to split objects (e.g., watches segment into dial and band, eggs into yolk and white). If replacing the segmenter with one that yields complete masks does not reduce CountZES's error on those classes, then exemplar selection itself is not what is driving the reported accuracy; the same could be tested by measuring whether selected exemplar masks cover at least 80% of the annotated instance.","supporting_citations":[],"review_version":1}