{"id":"c4d02ddc-ce0b-4ba6-b90a-8c7783ed3ed1","arxiv_id":"2506.01778","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"unMORE learns three object-centric representations (existence, center field, boundary distance) and uses a network-free reasoning algorithm to discover multiple objects in real-world images, outperforming prior unsupervised methods.","lead":"A new two-stage system, unMORE, finds and outlines many objects in single photos without any human-labeled training data. It first learns object existence, center, and boundary signals from ImageNet, then uses a hand-designed search procedure to discover objects in crowded real-world scenes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central SOTA claim is not yet established: hyperparameters are tuned on the same COCO* validation set used for reporting, COCO20K is not zero-shot for a COCO-trained detector, and no variance estimates are given.","rationale":"The reader's designated weakest assumption—that VoteCut pseudo masks in Sec 3.1 are a reliable proxy for object location and shape—is real, but it is partially mitigated by the rough-mask ablation in App. A.12 (Table 13): swapping VoteCut for MaskCut or SelfMask changes unMOREdisc APbox50 by only about 3 points, and even supervised VoteCut+SAM improves it by only 2.8 points. This suggests the pipeline is not tightly coupled to VoteCut specifically, so the pseudo-mask concern, while valid, is not the most load-bearing. The evaluation-protocol issues, by contrast, are unaddressed anywhere in the paper and directly gate the 'significantly outperforms' claim: thresholds tuned on the reporting set, a self-created primary benchmark, a COCO20K evaluation that is not zero-shot for a COCO-trained detector, and absent variance estimates. A conditional acceptance requiring a held-out tuning split, multiple seeds, and a corrected COCO20K protocol is the right call. I therefore keep the reader's CONDITIONAL verdict unchanged. A secondary reproducibility note: the prose in Sec 3.3 says the center-field kernel 'points outward,' while App. A.3 says it points 'towards the center' and Eq. (1) defines center-field vectors pointing away from the object center; the code release should clarify which convention is used, since an implementer following the prose alone could build a non-functional reasoning module.","tokens_in":40245,"tokens_out":18219,"duration_ms":191384,"concrete_test":"Hold out a random 20% of COCO* val for all hyperparameter selection (tau_e, tau_c, tau_conf, update margins), then freeze those settings and evaluate unMOREdisc and unMORE on the remaining 80%, repeating detector training with 3 seeds and reporting mean +- std. Also train an ImageNet-only variant of unMORE's detector (without COCO train pseudo labels) and compare it to CutLER Setting #3 on COCO20K. If the margin over CuVLER/CutLER falls within seed-level noise, or if the ImageNet-only variant no longer beats CutLER on COCO20K, the headline SOTA claim should be downgraded to 'comparable with favorable tuning.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim—significantly outperforming all existing unsupervised methods on 6 real-world benchmarks, especially COCO—depends on the validity of its evaluation protocol, and that protocol is currently insecure in three ways. First, all unMORE hyperparameters (existence threshold tau_e, center threshold tau_c, pseudo-label thresholds tau_conf in App. A.5/A.12, and the convergence margin in App. A.3) are selected using the same COCO* validation set that is reported as the headline benchmark, while baseline settings are taken from their original papers; this makes the comparison asymmetric. Second, COCO* is a manually augmented dataset created by the authors (App. A.16), with 10,336 additional labels produced through the authors' notion of 'objects of interest' using SAM; no independent group has validated these annotations, and the benchmark may favor a method explicitly designed to discover exactly such objects. Third, the zero-shot result on COCO20K is not zero-shot for unMORE: the best detector is trained on COCO 2017 train pseudo labels, and COCO20K is a subset of COCO 2014 trainval, which is contained in COCO 2017 train. CuVLER shares this overlap, but CutLER's best model is ImageNet-only, so the comparison to CutLER on COCO20K is unfair. Finally, no error bars or multiple-seed variance are reported anywhere, so the headline margins (e.g., +4.6 APbox50 over CuVLER in Table 1) have no stated uncertainty. These issues directly condition the claim that unMORE 'significantly outperforms all existing unsupervised methods,' and none of them is addressed in the manuscript.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes unMORE, a two-stage unsupervised multi-object segmentation pipeline. In the first stage, an objectness network is trained on ImageNet using VoteCut pseudo masks to predict three object-centric representations: a binary object existence score, an object center field, and an object boundary distance field. In the second stage, a network-free multi-object reasoning module initializes box proposals, filters them by existence scores, splits them by multi-center detection, and refines them by boundary-distance reasoning; optionally, the discovered objects are used as pseudo labels to train a class-agnostic detector. The paper claims state-of-the-art results on COCO* and several other real-world datasets, with particular strength in crowded images.","tokens_in":40590,"tokens_out":8268,"duration_ms":84621,"significance":"If the empirical claims are validated, unMORE would be a meaningful advance in unsupervised multi-object segmentation: it replaces pure feature-grouping with explicit learned object representations and a transparent, network-free reasoning procedure, and the mathematical property underlying the boundary update (Eq. 4) is correct. The paper also contributes a large COCO* re-annotation and promises code/data release. However, the central SOTA claim is currently conditioned on evaluation-protocol issues: hyperparameters are selected on the same COCO* validation set as the final results, the COCO20K 'zero-shot' experiment is not zero-shot for COCO-trained detectors, and no variance estimates are reported. These issues are fixable and do not appear to indicate a fundamental flaw in the method, but they must be addressed before the headline claims can be considered established.","major_comments":[{"comment":"The headline result on COCO* val is not as clean as presented because hyperparameters are tuned on that same set. Appendix A.12 reports selection of tau_e_conf, tau_c_conf, tau_b_conf on the COCO* validation set (Table 10) and selection of the binary-baseline step size on a 100-image subset of COCO* val (Table 9), while Table 1 reports the final numbers on the full COCO* val set; the baselines are taken at their original-paper settings without equivalent tuning. This makes the comparison asymmetric and weakens the statement in Sec. 4 that 'all final evaluation is conducted on COCO* val set which is completely held out.' I request an independent tuning split with the final table computed on the untouched COCO* val, or a sensitivity analysis demonstrating that the reported margins are stable over a range of thresholds, applied to the baselines as well.","section":"§4.1 / App. A.12 / Table 1"},{"comment":"The 'zero-shot' result on COCO20K is not zero-shot for unMORE (or for CuVLER in its best setting). According to App. A.9, the best unMORE detector (Setting #2) is trained on pseudo labels from the COCO 2017 train set, and CuVLER's best Setting #4 also uses COCO train pseudo labels, while CutLER's best Setting #3 is ImageNet-only. COCO20K is, per App. A.7, a subset of COCO 2014 trainval, from which the COCO 2017 train images are drawn; the COCO20K test images are therefore not held out for the COCO-trained detectors. The comparison against CutLER is thus unfair, and the phrase 'zero-shot detection' is misleading. Please report an ImageNet-only unMORE detector on COCO20K, or explicitly reframe this part of Table 2 as in-domain generalization and compare only with detectors having the same training-data overlap.","section":"§4.2 / App. A.9 / Table 2"},{"comment":"No error bars or multiple-seed variance are reported for any of the headline numbers. Proposal generation and detector training involve stochastic sampling (App. A.3, A.6), so margins such as +4.6 APbox50 over CuVLER on COCO* val in Table 1 and +1.8 APbox50 over CuVLER on COCO20K in Table 2 could lie within run-to-run variability. I ask the authors to report mean plus/minus standard deviation over at least three random seeds for the main tables, or to provide a formal significance test, for at least unMORE and the strongest baselines.","section":"§4.1 / Table 1 / §4.2 / Table 2"},{"comment":"The primary evaluation benchmark, COCO*, is an author-created re-annotation of COCO val2017 with 10,336 additional SAM-assisted labels across 197 new categories. There is no independent validation of these annotations, such as inter-annotator agreement statistics or a detailed public annotation protocol. Because the method is explicitly designed to discover precisely the kinds of objects that the authors chose to label, this benchmark may favor unMORE in a way that is not yet demonstrated. The manuscript should report the full annotation protocol, provide agreement statistics on a subsample, and give the original COCO val results (currently in App. A.10) equal prominence in the main text.","section":"§4 / App. A.16"}],"minor_comments":[{"comment":"Equation (9) in the appendix appears to have a sign typo: 'P_u1 <- -P_u1 - ...' should likely read 'P_u1 <- P_u1 - ...' to match the corresponding update in Eq. (6) of the main text.","section":"App. A.3, Eq. (9)"},{"comment":"The abstract states '6 real-world benchmark datasets,' but Section 4 lists seven named datasets (COCO*, COCO20K, LVIS, VOC, KITTI, Object365, OpenImages) in addition to the original COCO val in the appendix; please clarify the intended count.","section":"Abstract / Sec. 4"},{"comment":"Table 15 contains several spelling and truncation errors ('pinapple', 'spagatti', 'guita', 'envolop', 'vegatable', 'applicance', 'tomato 53' ordering), which should be corrected in the released annotation metadata as well as in the paper.","section":"App. A.16, Table 15"}],"recommendation":"major_revision","confidential_remarks":"The method itself appears coherent and the Eq. (4) derivation is sound, so the main risk is empirical rather than conceptual. I would condition acceptance on a re-evaluation that separates hyperparameter tuning from the test set, reports a true zero-shot COCO20K comparison, and includes variance estimates. If the authors can produce those, the contribution could be suitable for publication; the COCO* annotation release is also valuable if independently validated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one. The method is a real departure from the feature-clustering line: train an objectness network on ImageNet to predict three explicit representations (existence, center field, boundary distance field), then run a network-free iterative reasoning loop that splits proposals at anti-center peaks and expands/contracts boxes using the boundary field. The center-boundary reasoning is simple and effective, especially on crowded images, and the ablations are properly done—each component earns its keep. The paper is thorough and the math checks out.\n\nThe problems are all in the evaluation. First, thresholds for pseudo-label selection and some reasoning parameters are tuned on COCO* val, and COCO* val is also the headline benchmark. That makes the comparison to baselines (which use published settings) asymmetric. Not fatal, but the +4.6 AP over CuVLER is not as clean as it looks. Second, COCO* is the authors' own manually expanded annotation of COCO val. They add ~10k objects with SAM-assisted labeling. No independent check, and a method designed to discover 'objects of interest' may be favored. They do report original COCO val numbers, and those still show gains, so this is a caution, not a disqualifier. Third, and this is the real problem: the COCO20K 'zero-shot' claim is wrong. unMORE trains a detector on COCO 2017 train pseudo labels, and COCO20K is a subset of 2014 trainval that is inside that set. CutLER's best detector is ImageNet-only, so the comparison to CutLER on COCO20K is not a zero-shot comparison. That table needs to be rethought. Fourth, no error bars or seed variations anywhere; a 4.6 AP margin without variance is uninterpretable.\n\nThe 'all baselines collapse' phrase is also too strong. On crowded images, baselines degrade sharply, but unMORE's own AP also drops compared to sparse images. The relative gain is real, but 'collapse' oversells it.\n\nBottom line: the core idea deserves attention, and the paper is well-executed. The empirical claims need a careful redo before they can be trusted. I'd send it to referees—this is exactly the kind of paper where a good reviewer can separate the method from the overclaims. I would not cite the current numbers, but I would cite the mechanism once it's cleaned up.","headline":"A genuinely new mechanism for unsupervised multi-object segmentation, but the reported SOTA margins are not yet credible given the evaluation protocol.","tokens_in":41152,"tokens_out":4411,"would_cite":true,"duration_ms":45541,"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":"Explicit center and boundary fields, followed by network-free reasoning, let a single image yield many discovered objects without human labels, outperforming all unsupervised baselines on six benchmarks.","keywords":["unsupervised object segmentation","object-centric representation","center field","boundary distance field","network-free reasoning","multi-object discovery","crowded scenes","pseudo-label training"],"falsifier":"Take a test set of crowded images where the VoteCut pseudo masks demonstrably merge adjacent instances, run unMORE on them, and count whether the center-field splitting step separates those pairs; if unMORE's separations coincide with boundaries already present in the pseudo masks, the claimed gain over feature-grouping baselines reduces to pseudo-mask quality, while if it separates pairs the pseudo masks merged, the reasoning itself is doing genuine object recovery. A second decisive check is the paper's own reported failure on overlapping objects with similar textures: modifying the boundary field to incorporate texture and observing whether that failure disappears would test whether boundary geometry alone carries the claim.","tokens_in":40012,"feed_emoji":"🖼️","tokens_out":8826,"duration_ms":89247,"temperature":0.7,"pith_summary":"The paper tries to establish that a machine can discover many objects in a single real-world image without human labels by explicitly learning three object-centric representations on single-object images and then reasoning over them with a handcrafted procedure. The method trains an objectness network on ImageNet to predict whether an object is present, a center field that points from each object pixel to the object center, and a boundary distance field that gives the signed distance from each pixel to the nearest boundary. A network-free reasoning module then queries this frozen network on cropped patches of a scene image, discarding patches without objects, splitting patches that contain multiple centers, and expanding or contracting boxes until each converges on one object. On six real-world benchmarks including COCO, the reported results surpass all unsupervised baselines, with the largest gains on crowded images where other methods merge several objects into one. If correct, this shows that explicit center and boundary cues can substitute for the weak supervision that reconstruction and feature-grouping objectives provide.","feed_headline":"Center-boundary reasoning segments many objects per image, no labels","feed_subtitle":"On six real-world benchmarks including COCO, the network-free loop beats every unsupervised baseline.","key_machinery":"The central object is the objectness network plus its network-free reasoning loop. The network is trained to output three representations: a scalar existence score; a center field in which every pixel inside an object carries a unit vector pointing toward that object's bounding-box center; and a boundary distance field in which each pixel carries a normalized signed distance to the nearest boundary, positive inside the object and negative outside. The boundary field has a property that makes it load-bearing: the gradient norm at any interior pixel recovers the maximum distance from the boundary to the innermost pixel, so the reasoning module can convert boundary distances into physical pixel movements. The reasoning module seeds proposals from anchor boxes, removes patches with low existence scores, splits patches whose center fields produce a high anti-center response under a fixed outward-pointing kernel, and iteratively moves each proposal's four borders inward or outward according to the boundary field until convergence. Final masks are read off by thresholding the center and boundary fields, and duplicate boxes are removed by non-maximum suppression.","core_discovery":"The central claim is that three levels of explicitly defined object-centric representations, learned from single-object images, are sufficient to discover many objects in complex scenes without any human annotation or additional trainable modules. The paper argues that the right representations are a binary object existence score, a unit-vector center field pointing to the object center, and a normalized signed boundary distance field, and that these three fields together let a frozen network answer where objects are, how many are present, and how far each object extends. The subsequent reasoning module is network-free: it generates anchor proposals, filters them by existence score, splits proposals whose center fields show multiple centers, and refines each proposal's borders from boundary-distance values until the bounding box is tight. The paper reports that this pipeline outperforms all existing unsupervised methods on six real-world benchmarks, and that the improvement is most pronounced on crowded images, where feature-grouping baselines under-segment by grouping adjacent objects into one.","pith_inferences":["Because the reasoning module reads only geometric fields from a frozen network, the same center and boundary representations could be reused for interactive segmentation or point-and-click object editing, where the center field supplies the object anchor and the boundary field supplies the extent.","The COCO* re-annotation effort, which adds 197 categories and many previously unlabelled objects, may shift how unsupervised discovery methods are evaluated, since standard COCO annotations undercount objects and can turn correct discoveries into apparent false positives.","A testable extension is to train the objectness network from multiple different pseudo-mask generators and compare discovery quality; if unMORE's advantage shrinks when the pseudo masks are already excellent, most of the gain comes from supervision quality, whereas if it persists, the center-boundary reasoning is doing the work.","The paper's reported failure on overlapping objects with similar textures points to adding a texture-disambiguation term to the boundary field as the next step, rather than simply scaling the network."],"forward_implications":["Because the reasoning module is network-free, future improvements to the objectness network can be dropped in without retraining the discovery procedure.","A class-agnostic detector trained on unMORE's discovered objects inherits the multi-object recall and, the paper reports, beats detectors trained on earlier pseudo-label pipelines on COCO* and six zero-shot datasets.","Crowded images, where feature-grouping baselines merge adjacent objects into one, become tractable because the center-field splitting step is designed specifically to separate multiple centers inside one proposal.","The boundary distance field carries most of the performance gain in the ablations, suggesting that explicit boundary supervision, rather than a binary mask or reconstruction loss, is what lets the reasoning module tighten boxes accurately."],"supporting_citations":[{"why":"Supplies the VoteCut pseudo-mask procedure on ImageNet that generates the rough object masks used as training targets for the objectness network, and serves as a main baseline.","marker":"Arica et al., 2024"},{"why":"Supplies the self-supervised DINO features whose patch affinities VoteCut turns into object masks.","marker":"Caron et al., 2021"},{"why":"Supplies DINOv2 features, the stronger feature set used by VoteCut in the main experimental configuration.","marker":"Oquab et al., 2023"},{"why":"Provides the Normalized Cut algorithm that VoteCut uses to transform feature affinities into object masks.","marker":"Shi and Malik, 2000"},{"why":"Provides the Faster R-CNN anchor generation scheme that seeds the initial bounding-box proposals in the multi-object reasoning module.","marker":"Ren et al., 2015"},{"why":"Introduces the signed-distance and boundary-distance field formulation that the object boundary distance field is based on.","marker":"Park et al., 2019"},{"why":"Presents CutLER, the pseudo-label detector-training pipeline that unMORE extends and the main baseline it outperforms.","marker":"Wang et al., 2023a"},{"why":"Provides DINOSAUR, the reconstruction-based object-centric baseline that motivates the comparison against image-reconstruction objectives.","marker":"Seitzer et al., 2023"}],"fun_headline_variants":["No labels, no network: center-field reasoning finds many objects","Unsupervised multi-object segmentation via center-boundary fields","Center-boundary reasoning: unsupervised multi-object discovery without labels","No labels, network-free reasoning segments crowded scenes into objects","Explicit center and boundary fields beat all unsupervised baselines on COCO"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that VoteCut's pseudo masks on ImageNet mark objects reliably enough to teach the objectness network real objectness; if those masks systematically merge adjacent objects, miss small ones, or drag in background, the learned center and boundary fields inherit those errors and the network-free reasoning cannot recover objects the pseudo masks never represented.","fun_headline_variants_meta":{"raw":{"variants":["No labels, no network: center-field reasoning finds many objects","Unsupervised multi-object segmentation via center-boundary fields","Center-boundary reasoning: unsupervised multi-object discovery without labels","No labels, network-free reasoning segments crowded scenes into objects","Explicit center and boundary fields beat all unsupervised baselines on COCO"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00061,"raw_usage":{"total_tokens":2817,"prompt_tokens":897,"completion_tokens":1920,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":1835}},"tokens_in":513,"tokens_out":1920,"duration_ms":12037,"temperature":1.0,"reasoning_tokens":1835,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:33:40.497690+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a test set of crowded images where the VoteCut pseudo masks demonstrably merge adjacent instances, run unMORE on them, and count whether the center-field splitting step separates those pairs; if unMORE's separations coincide with boundaries already present in the pseudo masks, the claimed gain over feature-grouping baselines reduces to pseudo-mask quality, while if it separates pairs the pseudo masks merged, the reasoning itself is doing genuine object recovery. A second decisive check is the paper's own reported failure on overlapping objects with similar textures: modifying the boundary field to incorporate texture and observing whether that failure disappears would test whether boundary geometry alone carries the claim.","supporting_citations":[],"review_version":1}