{"id":"92d3318b-a8d6-4a02-9542-bd2583308b12","arxiv_id":"2505.02071","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"COCA-Net introduces compactness-guided hierarchical clustering within an attention architecture, achieving state-of-the-art unsupervised object segmentation on synthetic multi-object images.","lead":"A new attention-based clustering layer, COCA, uses a physical compactness score to pick object centroids and builds a hierarchy that segments multi-object scenes without labels. The resulting COCA-Net beats or matches existing unsupervised object-centric models on six synthetic benchmarks, and also produces usable segmentation masks directly from its encoder.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sequential clustering masks compactness scores with the scope but does not recompute affinity masks or compactness on the residual, so the claimed SBC concealment does not actually govern subsequent anchor selection.","rationale":"The reader's weakest assumption concerns the compactness prior itself, and that is a reasonable concern about generalization to non-convex or elongated objects. However, the more immediate technical gap is that even granting the prior, the sequential clustering procedure never re-scores the residual scene: Eq. 5a only erodes already-computed compactness scores, leaving the affinity maps and their contributions intact. The paper's own description says concealment should ensure only unassigned nodes contribute, but for anchor-selection scores that is not what the algorithm does. This is a concrete, localized, and testable issue in the core mechanism. It is not fatal: the empirical results across six synthetic datasets are extensive, the RAS ablation supports compactness as a useful signal, and end-to-end training can partially adapt features to the scoring procedure. The reader's additional conditions about dynamic-slot overclaiming and unspecified baseline encoder-mask computation also remain valid. Because the central SOTA claim is not directly refuted but is less secure than stated, the CONDITIONAL verdict should stand, now including a request to report residual-recomputation behavior.","tokens_in":23975,"tokens_out":11346,"duration_ms":143391,"concrete_test":"Implement a residual-recomputation variant of Algorithm 1: after each scope update Z_m, recompute affinity maps Λ and compactness scores C only over unassigned nodes (equivalently, mask Λ by Z_m before Eq. 3), keeping all other hyperparameters fixed. Train on CLEVR6 and evaluate on CLEVR6 and CLEVR10, reporting decoder foreground ARI/mSC over the same three seeds. If scores change by more than about 2-3 points, the one-shot compactness approximation is material; if they do not, the incomplete residual masking is harmless in practice. Also compare anchor node positions against ground-truth object centroids on a subset of images with overlapping objects.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1 and Eqs. 5a-5d select anchors by compactness scores C_l, but C_l is computed once per layer from the full affinity maps Λ_l (Eq. 3), before sequential clustering begins. At iteration m, Eq. 5a only multiplies C_l by the scope Z_{m-1}; it does not recompute Λ_l or C_l on the residual set of unassigned nodes. Consequently, a remaining node's compactness score still includes contributions from pixels that have already been assigned to earlier clusters. The text in Sec. 3.1.5 and Algorithm 1 line 3 says concealment ensures 'only unassigned nodes contribute to subsequent calculations'; this is not true for the anchor-selection signal. For overlapping or similar-looking objects, the true centroid of an unassigned object can be penalized by its affinity to already-concealed pixels, so anchor selection may miss it or select a subregion. Because Eq. 3 compactness is the only inductive bias for choosing object centroids, this incomplete residual masking is load-bearing for the claimed SBC-based clustering behavior.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Compact Clustering Attention (COCA) layer, an attention-based clustering module for unsupervised object-centric learning. COCA scores candidate affinity masks by a moment-of-inertia compactness measure and sequentially selects the most compact mask as an object cluster using a stick-breaking concealment strategy. Stacking these layers forms COCA-Net, which is trained end-to-end with a spatial broadcast decoder. The authors evaluate on six synthetic datasets (Tetrominoes, Multi-dSprites, ObjectsRoom, ShapeStacks, CLEVR6, CLEVRTex) against GEN-v2, INVSA, and BOQSA, reporting decoder and encoder segmentation masks, with and without background, and include ablations for compactness-based versus random anchor selection and for dynamic slot allocation.","tokens_in":24255,"tokens_out":5522,"duration_ms":65612,"significance":"If the results hold, COCA-Net offers a credible alternative to slot-attention models, with particular strengths in encoder-side segmentation, background handling, and lower variance across training runs. The compactness functional is an external geometric measure from the cited literature, not fitted to the data, and evaluation is on held-out test images, so circularity concerns are limited. The ablations in Table 3 and supplementary Tables 9-10 provide machine-checked support for the central design choice of compactness-guided anchor selection. However, the conceptual novelty rests on the claimed stick-breaking concealment mechanism and on the compactness prior; both need closer scrutiny, as detailed below.","major_comments":[{"comment":"The compactness scores C^l are computed once per layer from the full affinity masks Λ^l (Eq. 3), before sequential clustering begins. Eq. (5a) only multiplies C_{m-1} by the scope Z_{m-1}; it does not recompute Λ^l or C^l on the residual set of unassigned nodes. Consequently, the sentence in Sec. 3.1.5 that this \"ensuring that only unassigned nodes (with non-zero scope values) contribute to subsequent calculations\" is inaccurate for the anchor-selection signal. A remaining node's compactness score still includes contributions from pixels already assigned to earlier clusters, so the selected anchor may not be the most compact residual shape. This issue is load-bearing because compactness is the only inductive bias for choosing object centroids. The authors should either modify the algorithm to recompute compactness after masking the affinity masks by the scope, or revise the text to state that scores are computed once per layer and provide an experiment or argument demonstrating that stale scores do not degrade anchor selection.","section":"3.1.5, Eqs. (5a)-(5d) and Algorithm 1"},{"comment":"The paper assumes, in Sec. 3.1.5, that nodes corresponding to the centroids of distinct objects yield the highest compactness scores within their objects, and that foreground objects are more compact than background regions. This assumption is validated only on synthetic datasets composed of convex, blob-like objects. For elongated, concave, or hollow objects, or when background regions are more compact than foreground, the anchor-selection mechanism would likely fail. Because this compactness prior is the central spatial inductive bias of the method, the authors should either restrict their claims to the considered benchmark, or include a stress test with non-convex shapes (e.g., synthetic objects with holes or elongated structures) to characterize the limits of the approach.","section":"3.1.5, compactness assumption"}],"minor_comments":[{"comment":"The abstract and the concluding paragraph of Sec. 4.3 claim that COCA-Net is \"not bound by a predetermined number of object masks,\" but all main experiments fix the number of output slots to the dataset maximum, and only the COCA-Net-Dyna ablation supports the dynamic-slot claim. Please qualify this claim in the abstract and conclusion.","section":"Abstract and Sec. 4.3"},{"comment":"In Algorithm 1, line 6 of the supplementary material, the scope update is written as \"Zl = Zl ⊙ (1− Πl−1)\"; this appears to use an undefined Πl−1. It should be the current cluster mask Πm (or Π appended in the previous line). Please correct the notation.","section":"Supplementary, Algorithm 1, line 6"},{"comment":"Table 3 and the supplementary Tables 9-10 report single-seed results, whereas Tables 1 and 2 report mean ± standard deviation over three seeds. To avoid confusion when comparing these results, the captions should state explicitly that these are single-seed numbers.","section":"Tables 3, 9, 10"},{"comment":"Equation (3) includes a summation over j < v without defining the index v in the main text. Please define all summation indices and the range of v.","section":"Eq. (3)"},{"comment":"The text states that \"A perfect circle achieves the maximum compactness value of 1,\" but Eq. (3) evaluates compactness around an arbitrary node i; the maximum may be less than 1 when i is not the centroid of the shape. Please rephrase to clarify that the bound applies when the reference point is the centroid.","section":"Sec. 3.1.4"}],"recommendation":"major_revision","confidential_remarks":"The empirical results are interesting and the paper is generally well written, but the main technical concern in my report—the discrepancy between the described stick-breaking concealment and the actual implementation—should be addressed before acceptance. The authors also do not provide a code release link in the manuscript; given that the method is a new architecture, releasing code would substantially strengthen reproducibility and the ability of reviewers to verify implementation details."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nCOCA-Net is worth a careful read. It is a genuinely new architecture that brings hierarchical clustering and a moment-of-inertia compactness prior into unsupervised object-centric learning, and it backs the central design choice with a clean ablation. The encoder-side dendrogram masks and the improved background segmentation are real contributions, and the evaluation across six synthetic datasets with three-seed variance reporting is above the usual bar.\n\nWhat impressed me: the compactness measure comes from the geography literature, is not fitted to the data, and the ablation in Table 3 (compactness vs. random anchor selection, ARI 0.894 vs. 0.832 on ObjectsRoom) supports the core claim. The improvement in encoder-side masks over Slot-Attention baselines is large and consistent. The authors also list limitations honestly in the conclusion.\n\nSoft spots, in proportion:\n\nFirst, the abstract overclaims. It says COCA-Net is \"not bound by a predetermined number of object masks,\" but the main experiments fix the number of output slots to the dataset maximum, as stated in Section 4.1. Dynamic allocation only appears in the Dyna ablation. That claim should be tempered.\n\nSecond, the stress-test concern about Algorithm 1 is valid. The compactness scores C_l are computed once per layer from the full affinity maps, before sequential clustering begins. Equation 5a only erodes the scores by the scope; it does not recompute C_l on the residual set. So a remaining node's compactness score still contains contributions from already-assigned pixels. The text's statement that \"only unassigned nodes contribute to subsequent calculations\" is therefore inaccurate. On the tested synthetic scenes with mostly non-overlapping blobs this may not hurt, but it could break anchor selection for overlapping or concave objects. The authors should either correct the description or actually recompute compactness on the residual.\n\nThird, the compactness prior is a domain assumption: it expects objects to be compact, convex blobs. The paper only validates on synthetic datasets; the real-world experiments are preliminary, single-seed, and use a different decoder. That is acceptable as a limitation but should be stated plainly.\n\nFourth, baseline encoder masks are not clearly described. For peer review, the authors need to explain how encoder-side masks were obtained for GEN-v2, INVSA, and BOQSA. Code or checkpoints would also help; none are linked.\n\nBottom line: the paper deserves a serious referee. The core idea is sound and the ablation supports it. The concealment flaw is a real but fixable accuracy issue in the description, not a collapse of the method. I would send it to review, and ask for a major revision addressing the overclaim, the residual-compactness issue, and the reproducibility details.\n\nRecommendation: engage with it. Bring it to the reading group if you care about object-centric learning; there is a useful discussion here about what \"concealment\" actually buys you.","headline":"COCA-Net is a genuinely new compactness-guided hierarchical clustering architecture for object-centric learning with strong synthetic results, but the abstract overclaims dynamic slots and the sequential clustering concealment is not what it says it is.","tokens_in":24749,"tokens_out":3634,"would_cite":true,"duration_ms":41302,"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":"Compactness, not preset slots, drives this unsupervised object segmenter.","keywords":["object-centric learning","unsupervised segmentation","compactness prior","moment of inertia","hierarchical clustering","slot attention","sequential clustering","scene decomposition"],"falsifier":"Run COCA-Net unsupervised on a synthetic dataset of hollow rings, elongated rods, or U-shaped objects: if the highest-scoring anchor consistently lands away from the object's perceptual center, or on the background, the central premise fails and segmentation accuracy should drop sharply relative to the convex-blob datasets. The check can be made directly by comparing the argmax of Eq. 3 to the true object centroids on such shapes.","tokens_in":23804,"feed_emoji":"🎯","tokens_out":4399,"duration_ms":45066,"temperature":0.7,"pith_summary":"The paper proposes COCA-Net, an unsupervised architecture that discovers objects in an image by clustering pixels hierarchically and, at every step, asking which candidate region is most compact. Its central claim is that a physical notion of compactness, the ratio of a shape's moment of inertia to that of a circle of equal area, identifies the centroid of a distinct object, so that repeatedly selecting the most compact remaining region and concealing it yields one object mask per iteration. The authors argue this spatial inductive bias fixes the main weaknesses of Slot-Attention models: sensitivity to initialization, a hard requirement to preset the number of slots, and poor background segmentation. Across six synthetic datasets and nine metrics, they report that COCA-Net outperforms or matches state-of-the-art baselines, with notably lower variance across training runs and strong masks coming directly from the encoder. If true, this means object-centric learning can rest on an explicit, interpretable notion of objectness rather than on learned slot routing.","feed_headline":"A compactness score, not preset slots, drives object discovery","feed_subtitle":"New hierarchical clustering layer beats or matches slot-attention models on six benchmarks, with no fixed object count.","key_machinery":"The load-bearing object is the compactness functional $C_i(\\Lambda^l_i)$ of Eq. 3, which measures a candidate affinity mask's moment of inertia around node $i$ against that of a disk of equal area; because inertia is minimized at the true centroid, the node at an object's centroid receives the highest compactness score within that object. This single score drives anchor selection in the sequential stick-breaking clustering loop (Eqs. 5a-5d), which in turn determines all cluster assignments in the hierarchy.","core_discovery":"COCA-Net claims that unsupervised object discovery reduces to a greedy search over compactness. A COCA layer refines pixel features with self-attention, computes an affinity mask for every node in a window, scores each mask with a mass-normalized moment-of-inertia compactness functional (Eq. 3), and then runs a stick-breaking concealing loop (Eqs. 5a-5d): take the highest-scoring remaining mask as the next cluster, mask out its nodes, and repeat until the scope is exhausted. Stacked in a bottom-up hierarchy over non-overlapping windows, this produces a dendrogram of object masks with a variable number of clusters, trained end-to-end with only a pixel-reconstruction loss. The paper reports state-of-the-art or competitive results on Tetrominoes, Multi-dSprites, ObjectsRoom, ShapeStacks, CLEVR6, and CLEVRTex across ARI and mean Segmentation Covering, on both the decoder side and, unusually, the encoder side, with background regions segmented more coherently than by the baselines.","pith_inferences":["A sharp empirical boundary follows from the compactness prior: the method should transfer to real-world images only insofar as objects are convex and blob-like; on elongated, hollow, or heavily concave objects the Eq. 3 anchor selection would likely drift from the perceptual center, a regime the six synthetic benchmarks do not probe.","Because the dendrogram is produced deterministically from compactness scores, encoder-side masks could be emitted without a decoder forward pass, suggesting a cheap inference mode for downstream tasks that the paper hints at but does not test.","The explicit geometric prior yields a falsifiable prediction that slot-attention models lack: deforming an object (stretching it, adding holes, making it concave) should degrade COCA-Net's anchor selection in a monotone, predictable way, and this degradation could be measured directly on synthetic shape morphs."],"forward_implications":["Dynamic slot allocation: because clustering stops when the scope is exhausted or drops below a threshold, the number of output masks is not fixed in advance; the COCA-Net-Dyna variant trained on CLEVR6 transfers to CLEVR10 with only a small drop (ARI 0.978 vs 0.985).","The encoder alone produces segmentation masks that match or beat decoder masks on most datasets, so the trained hierarchy could be reused as a standalone unsupervised object-centric feature extractor.","Background elements are segmented as coherent clusters: with background included in the evaluation, COCA-Net gains roughly thirty ARI points over INV-SA and BOQ-SA on ObjectsRoom, where those baselines collapse to background ARI around 0.6 or lower.","Training is more stable: over three seeds COCA-Net shows markedly smaller standard deviation than GEN-v2, INV-SA, and BOQ-SA on nearly every dataset and metric.","The windowed hierarchy operates on $U\\times U$ windows in parallel with per-layer complexity reducible to $O(N^2\\log N)$, making the architecture amenable to parallel implementation and scaling to higher resolutions."],"supporting_citations":[{"why":"Defines Slot Attention, the baseline paradigm and K-Means-like clustering analogy that COCA argues against and is compared with.","marker":"[38]"},{"why":"Supplies the sequential stick-breaking concealing strategy (SBC) that COCA adapts, replacing random anchor sampling with compactness-based selection.","marker":"[18]"},{"why":"One of the three state-of-the-art baselines; its invariant-slot reference-frame design is compared on all six datasets.","marker":"[5]"},{"why":"One of the three baselines; the query-optimized slot attention variant that COCA-Net must match or beat.","marker":"[28]"},{"why":"Provides the benchmark library, the six datasets, and the training and evaluation protocol used for all experiments.","marker":"[15]"},{"why":"Introduces the mass-normalized moment-of-inertia compactness measure from which Eq. 3 is derived, including the centroid-minimizing property.","marker":"[55]"},{"why":"Describes the Spatial Broadcast Decoder used to reconstruct images and produce the decoder-side segmentation masks.","marker":"[52]"},{"why":"Supplies the ViT-22B layer recipe (self-attention with parallel FFN and pre-norm skip connections) adapted for feature refinement in COCA layers.","marker":"[13]"}],"fun_headline_variants":["Compactness, not preset slots, guides object discovery in COCA-Net","COCA-Net uses compactness clustering for unsupervised object discovery","Hierarchical compactness clustering achieves unsupervised object discovery","Object discovery without preset slots: COCA-Net's compactness approach","Compactness-guided hierarchical clustering excels at object discovery"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that, within any window, the node whose affinity mask is most compact by the moment-of-inertia score is the centroid of a distinct foreground object, and that compactness separates objects from background; this premise is only validated on synthetic scenes of convex, blob-like objects.","fun_headline_variants_meta":{"raw":{"variants":["Compactness, not preset slots, guides object discovery in COCA-Net","COCA-Net uses compactness clustering for unsupervised object discovery","Hierarchical compactness clustering achieves unsupervised object discovery","Object discovery without preset slots: COCA-Net's compactness approach","Compactness-guided hierarchical clustering excels at object discovery"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000189,"raw_usage":{"total_tokens":1336,"prompt_tokens":943,"completion_tokens":393,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":308}},"tokens_in":559,"tokens_out":393,"duration_ms":3215,"temperature":1.0,"reasoning_tokens":308,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T01:02:47.193409+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run COCA-Net unsupervised on a synthetic dataset of hollow rings, elongated rods, or U-shaped objects: if the highest-scoring anchor consistently lands away from the object's perceptual center, or on the background, the central premise fails and segmentation accuracy should drop sharply relative to the convex-blob datasets. The check can be made directly by comparing the argmax of Eq. 3 to the true object centroids on such shapes.","supporting_citations":[{"cited_title":"Object- centric learning with slot attention","cited_arxiv_id":null,"evidence_quote":"Defines Slot Attention, the baseline paradigm and K-Means-like clustering analogy that COCA argues against and is compared with."},{"cited_title":"Genesis-v2: Inferring unordered object representations with- out iterative refinement","cited_arxiv_id":null,"evidence_quote":"Supplies the sequential stick-breaking concealing strategy (SBC) that COCA adapts, replacing random anchor sampling with compactness-based selection."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"One of the three state-of-the-art baselines; its invariant-slot reference-frame design is compared on all six datasets."},{"cited_title":"Improving object- centric learning with query optimization","cited_arxiv_id":null,"evidence_quote":"One of the three baselines; the query-optimized slot attention variant that COCA-Net must match or beat."},{"cited_title":"Dittadi, S","cited_arxiv_id":null,"evidence_quote":"Provides the benchmark library, the six datasets, and the training and evaluation protocol used for all experiments."},{"cited_title":"Goodchild Wenwen Li and Richard Church","cited_arxiv_id":null,"evidence_quote":"Introduces the mass-normalized moment-of-inertia compactness measure from which Eq. 3 is derived, including the centroid-minimizing property."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the ViT-22B layer recipe (self-attention with parallel FFN and pre-norm skip connections) adapted for feature refinement in COCA layers."}],"review_version":1}