{"id":"95f836a7-08d9-4315-a619-edb4c293f6be","arxiv_id":"2502.06460","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A CLIP-based group re-identification method that simulates member changes with Bernoulli masks and learns uncertain text descriptions, reporting top Rank-1 accuracy on CSG, RoadGroup, and iLIDS-MCTS.","lead":"This paper adapts the CLIP image-text model to recognize groups of pedestrians across cameras, using learned text prompts that mention possibly present members to cope with changing group size and layout. The authors report state-of-the-art accuracy on three group re-identification benchmarks, though the evaluation protocol needs scrutiny.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim rests on an undocumented random 70/30 split: no identity-disjoint guarantee, seed, or variance, so Table I comparisons may be inflated and the central empirical claim is not verifiable as submitted.","rationale":"I read the paper in good faith. The architecture is coherent: MVS simulates member dropout via Bernoulli masks (Eq. 1), GLA builds group text templates with learnable tokens, and GRCE applies cross-attention to refine group features. The ablation in Table II shows monotonic gains from adding GLA, MVS, and GRCE, which is internally consistent. However, the paper's central claim is empirical SOTA performance, and that claim depends entirely on the evaluation being comparable to the baselines. The reader's weakest_assumption identifies exactly the fragile point: the random 70/30 split is not documented as identity-disjoint, and no seed or variance is reported. I agree. The margins over UMSOT (0.8-3.1 Rank-1 points) are small relative to typical protocol differences in ReID; without official splits or code, the table cannot be independently verified. I also note that no code, checkpoints, or formal verification are provided, so there is no independent support for the numerical results. This is not a complaint about the idea, which is plausible, but about the evidence required for the claim as stated. The appropriate disposition is therefore the same as the reader's: reject as submitted, with the path to acceptance being a matched-protocol, identity-disjoint evaluation with variance and code.","tokens_in":7853,"tokens_out":3352,"duration_ms":29687,"concrete_test":"Use the official identity-disjoint train/test partitions adopted by UMSOT and SOT on CSG, RoadGroup, and iLIDS-MCTS; retrain GCUM with the paper's hyperparameters on those partitions and report Rank-1 and mAP over at least 5 seeds as mean ± std. If, under a matched protocol, GCUM no longer beats UMSOT by more than 1.0 Rank-1 point on all three datasets, the 'significantly outperforms' claim fails. A cheaper minimal check: rerun only iLIDS-MCTS with an explicit identity-disjoint 70/30 split and 10 seeds; if the 67.8% vs 64.7% margin is not reproducible, the verdict should be reject.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's empirical claim—'significantly outperforms state-of-the-art' with Rank-1 accuracies of 94.4%, 90.1%, and 67.8% on CSG, RoadGroup, and iLIDS-MCTS (Section III-B, Table I)—requires that the reported numbers are produced under an evaluation protocol comparable to the protocols of the prior methods in Table I. Section III-A specifies only that each dataset is 'randomly divided' into 70% training and 30% test, with no statement that the split is identity-disjoint, no random seed, and no repeated runs. In Group ReID the retrieval unit is a group identity across cameras; if a random image-level or group-level split places some cameras of the same group identity in both training and test, the text descriptions and classifiers trained in Stage 2 (Eqs. 6-8) have already seen test identities, inflating Rank-1. The claimed margins over UMSOT are only 0.8 (CSG), 1.2 (RoadGroup), and 3.1 (iLIDS-MCTS) Rank-1 points, which is exactly the scale of protocol mismatch or a single unlucky draw. No code or checkpoints are provided. Thus the load-bearing condition—a comparable, identity-disjoint split with reported uncertainty—is unsupported, and the central claim is not verifiable from the submission.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes GCUM, a two-stage CLIP-based method for group re-identification. The first stage introduces a Member Variant Simulation (MVS) module and a Group Layout Adaptation (GLA) module to generate uncertain group text descriptions, while the second stage employs a Group Relationship Construction Encoder (GRCE) to refine group visual features and applies identity, triplet, and image-to-text cross-entropy losses. Experiments on iLIDS-MCTS, RoadGroup, and CSG report state-of-the-art Rank-1 accuracies of 67.8%, 90.1%, and 94.4%, and the paper claims to be the first to apply CLIP to group ReID.","tokens_in":8173,"tokens_out":5061,"duration_ms":42302,"significance":"If the reported results are reproducible under a correct evaluation protocol, the idea of using CLIP for group ReID with uncertainty modeling is novel and potentially impactful. Strengths include a concrete two-stage framework, ablations over the proposed modules, and retrieval visualizations. The empirical central claim, however, depends on an evaluation protocol that is underspecified in Section III-A, and the claimed generalization from text descriptions is not cleanly separated from a self-distillation loop. The margins over prior state-of-the-art are small, so evaluation rigor is critical.","major_comments":[{"comment":"The paper states only that each dataset is 'randomly divided' into a 70% training set and a 30% test set, without specifying that the split is identity-disjoint, without giving a random seed, and without reporting repeated runs. Since Table I compares against prior methods whose scores are typically reported under official identity-disjoint splits, this ambiguity makes the central SOTA claim unverifiable. If group identities appear in both training and test sets, the reported Rank-1 gains (0.8, 1.2, and 3.1 points over UMSOT) could stem entirely from protocol mismatch. Please specify the split policy, adopt the official splits if available, and report the mean and standard deviation over at least three seeds.","section":"Section III-A, Implementation details"},{"comment":"The claim of 'significantly outperforming' state-of-the-art is not supported by the reported numbers, since the Rank-1 margins over UMSOT are only 0.8 (CSG), 1.2 (RoadGroup), and 3.1 (iLIDS-MCTS) percentage points. Without variance estimates or statistical significance testing, these differences are within the range of random split variation. Additionally, the 'improvement' percentages over hand-crafted methods are computed as absolute percentage-point differences (e.g., 94.4 - 19.2 = 75.2) and are described as '% improvement,' which is misleading; please report relative improvements or explicitly state that these are percentage-point increases.","section":"Section III-B, Compared with Other Group Re-ID Methods"},{"comment":"The MVS module is underspecified. The notation 'p ∼ N (µ, σ; p0, pmax)' in Eq. (1) is not defined; it is unclear whether this is a truncated normal or a clipping operation, and the roles of p0 and pmax are not explained. In Eq. (2), the text says 'a represents selecting the first elements,' which is likely a typo for 'top'; the operation 'top(|− →m|)' is not defined elsewhere. Since MVS is a core contribution, these missing details prevent reproduction.","section":"Section II-A, Eqs. (1)-(2)"},{"comment":"The learnable text tokens [X] and [P] are optimized in Stage 1 on the same training identities that later supervise the visual features via the Li2tce loss in Stage 2. Therefore the 'generalizable knowledge from group text descriptions' is not externally grounded in CLIP's pretrained semantics but is a self-distillation loop on the training identities. The paper should clarify what information is transferred beyond the identity labels, or demonstrate generalization to unseen identities through a zero-shot or cross-dataset experiment.","section":"Section II-C, second training stage and Eq. (7)"}],"minor_comments":[{"comment":"The phrase 'adapts group text descriptions to undetermined accommodate member and layout variations' is ungrammatical; the word 'undetermined' appears to be an error and should be removed or replaced.","section":"Abstract"},{"comment":"The ablation text refers to a 'Dual-Branch' component, but the method section names this component 'GRCE'; please use consistent terminology throughout.","section":"Section III-C, Ablation studies"},{"comment":"The index 'K' in 'TyK' is inconsistent with the surrounding notation; it should likely be 'Tyi', and the denominator summation index should be clarified.","section":"Eq. (7)"},{"comment":"The mAP column is empty for several prior methods; please state whether mAP is unavailable for those methods and how the table should be interpreted.","section":"Table I"},{"comment":"The number of learnable tokens M and the predefined maximum group size K are not specified; please provide values for each dataset.","section":"Section II-C"},{"comment":"The caption does not explain the distinction between the Stage 1 and Stage 2 losses; a more detailed caption would improve readability.","section":"Figure 2 caption"}],"recommendation":"major_revision","confidential_remarks":"The central obstacle is the under-specified evaluation protocol: the random 70/30 split must be identity-disjoint for the Table I comparison to be meaningful, and the small margins over UMSOT make this a load-bearing issue. If the authors cannot confirm the split is identity-disjoint or provide code, the empirical claims should be withdrawn or substantially weakened. The self-distillation concern is significant but could be addressed with framing as a training strategy rather than a transfer of external knowledge. Overall, the paper is a candidate for revision rather than outright rejection because the issues are potentially fixable with additional experiments and clarity."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper's core idea—applying CLIP to group re-id with uncertainty text descriptions—is new and worth a look, but the reported SOTA numbers are not verifiable because the 70/30 split is undocumented. As submitted, the empirical claim doesn't hold.\n\nWhat's actually new: the paper is, as far as the cited literature goes, the first to adapt CLIP-ReID to group re-id. The combination of MVS (a Bernoulli mask to simulate member removal), GLA (learnable group tokens with identity-specific text), and GRCE (cross-attention to refine group features) is a sensible extension of CLIP-ReID. The idea of injecting 'potentially present' into the text template to handle count variation is a nice contribution conceptually.\n\nWhere it slips: the evaluation protocol is the load-bearing problem. Section III-A says 'randomly divide 70/30' without saying the split is identity-disjoint, without a seed, and without repeated runs. In group re-id, a group identity can appear across multiple cameras. If any camera of a test identity appears in training, the text descriptions and classifiers have already seen that identity, and Table I is inflated. The margins over UMSOT are 0.8, 1.2, and 3.1 Rank-1 points—exactly the scale of a protocol mismatch or an unlucky draw. No code or checkpoints are provided, so the numbers cannot be checked.\n\nThe method section also has under-specified pieces: Eq. (1) writes β(1−p) but the text says 0 with probability p, so the mask's distribution is misstated. Eq. (2) has an undefined 'a' and unclear top(k) notation. And the 'potentially present' mechanism—how MVS actually influences the text descriptions—is described only at a high level. The Stage-2 Li2tce uses text descriptions learned on the same training identities, so it's a self-distillation rather than transfer of externally grounded knowledge; that's fine as a regularizer, but the paper oversells the 'generalizable knowledge' part.\n\nThat said, these are fixable. The conceptual framework is coherent. Give it to a serious referee with a clear request for an identity-disjoint split, seeds, variance, code, and a rewritten MVS/GLA description. For readers in group re-id, it's a useful pointer for a CLIP-based direction, but Table I should not be trusted as-is.","headline":"New CLIP-based group re-id idea, but the SOTA claim rests on an undocumented 70/30 split and is not verifiable as submitted.","tokens_in":8684,"tokens_out":3059,"would_cite":false,"duration_ms":25336,"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":"The paper claims the first CLIP-based Group ReID method, using uncertain text descriptions to match groups whose member count and layout were never seen in training.","keywords":["Group Re-Identification","CLIP","uncertainty modeling","vision-language pre-training","member variant simulation","group layout adaptation","cross-modal contrastive learning"],"falsifier":"Retrain GCUM on an explicitly identity-disjoint split of CSG with a fixed seed, repeated for ten runs, and compare the mean Rank-1 to UMSOT's 93.6%; if the mean falls to or below that value, the claimed generalization advantage is not demonstrated.","tokens_in":7651,"feed_emoji":"👥","tokens_out":5159,"duration_ms":42227,"temperature":0.7,"pith_summary":"The paper tries to establish that a vision-language model can be made to match pedestrian groups across cameras even when the number of people in the group and their spatial layout change between views. Its proposal, GCUM, is claimed to be the first application of CLIP to Group Re-Identification. Instead of encoding fixed group structures, it learns uncertain text descriptions that include identity-specific tokens and phrases like 'potentially present,' and it simulates member dropout during training so those descriptions stay valid for unseen compositions. The method reports Rank-1 accuracies of 94.4% on CSG, 90.1% on RoadGroup, and 67.8% on iLIDS-MCTS, beating prior single-dataset state of the art by 0.8 to 3.1 points. If correct, this means group matching can be framed as a text-conditioned retrieval problem rather than a fixed-structure recognition problem.","feed_headline":"CLIP with uncertain text hits 94% on group re-ID","feed_subtitle":"Beats the prior best single-dataset method by up to 3.1 points by simulating missing members and shifting layouts.","key_machinery":"The Member Variant Simulation (MVS) module applies a Bernoulli mask to randomly remove members and re-weights the group class token with a learnable quantity-relationship matrix. The Group Layout Adaptation (GLA) module aggregates per-member learnable tokens into a group sentence with conditional options. The Group Relationship Construction Encoder (GRCE) refines individual features through cross-attention from group features and is trained with identity, triplet, and image-to-text contrastive losses. The load-bearing idea is that text descriptions can carry uncertainty explicitly where fixed visual templates cannot.","core_discovery":"GCUM is the first to bring CLIP to Group Re-ID, and it works by replacing fixed group-structure descriptions with uncertain text descriptions that include identity-specific tokens plus 'potentially present' phrasing. Training with simulated member exclusions and layout adaptation lets the text encoder align with visual features of groups whose member count and layout were never seen in training. On CSG, RoadGroup, and iLIDS-MCTS the method reports Rank-1 accuracies of 94.4%, 90.1%, and 67.8%, surpassing the previous single-dataset state of the art UMSOT by 0.8, 1.2, and 3.1 points respectively. These numbers, if they hold under a fair evaluation protocol, mean that uncertainty in text descriptions is a viable route to generalizing across group structures.","pith_inferences":["The 'potentially present' mechanism is effectively a form of compositional open-vocabulary learning; a natural test is whether the learned tokens can describe group sizes beyond the predefined maximum M0.","The same uncertainty-description recipe could be carried to other structured-object recognition tasks, such as matching vehicle convoys or animal herds, where membership is fluid.","The evaluation rests on a random 70/30 split; an identity-disjoint protocol with repeated seeds and standard deviations would likely shrink the reported margins but would be the fair test of the generalization claim.","One could add test-time member-mask augmentation to probe whether the model's robustness is truly continuous in the mask probability p."],"forward_implications":["A model trained on a single dataset can beat baselines that were trained on multiple datasets, because the text descriptions transfer structural knowledge rather than memorized layouts.","The MVS, GLA, and GRCE modules each contribute; ablating any of them drops Rank-1 by at least several points, so the method's gains are not from the backbone alone.","Retrieval visualizations show correct group matches even when member count and layout differ between query and gallery, indicating the learned text tokens are robust to structural variation.","The two-stage pipeline separates knowledge transfer from visual refinement, meaning either stage could be reused with a different backbone or a different text generator."],"supporting_citations":[{"why":"Introduces the multi-grain group-reID method LIMI that this work compares against and builds on.","marker":"[1]"},{"why":"Introduces the CSG dataset and the multi-attention context graph baseline MACG.","marker":"[2]"},{"why":"Supplies the CLIP-ReID backbone and the learnable-token text-template idea that GCUM extends.","marker":"[11]"},{"why":"Introduces the iLIDS-MCTS dataset and the earliest cross-camera group association protocol.","marker":"[16]"},{"why":"PREF, a hand-crafted baseline used to show the large margin of learned text descriptions.","marker":"[18]"},{"why":"The second-order-transformer uncertainty baseline SOT that the method must surpass.","marker":"[20]"},{"why":"UMSOT, the strongest single-dataset baseline; the reported gains are measured against its numbers.","marker":"[21]"}],"fun_headline_variants":["First CLIP-based group re-ID uses uncertain text to generalize","CLIP + uncertainty modeling boosts group re-ID to 94% on CSG","Simulated member loss and layout shifts improve CLIP group re-ID","Uncertain descriptions let CLIP handle unseen group structures","Group re-ID with CLIP: uncertainty beats fixed structure text"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains assume the random 70/30 train/test split keeps identities disjoint between partitions; the paper does not state a seed or confirm disjointness, so if identities overlap the accuracy numbers would be inflated.","fun_headline_variants_meta":{"raw":{"variants":["First CLIP-based group re-ID uses uncertain text to generalize","CLIP + uncertainty modeling boosts group re-ID to 94% on CSG","Simulated member loss and layout shifts improve CLIP group re-ID","Uncertain descriptions let CLIP handle unseen group structures","Group re-ID with CLIP: uncertainty beats fixed structure text"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000341,"raw_usage":{"total_tokens":1864,"prompt_tokens":919,"completion_tokens":945,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":535,"completion_tokens_details":{"reasoning_tokens":856}},"tokens_in":535,"tokens_out":945,"duration_ms":9131,"temperature":1.0,"reasoning_tokens":856,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T15:21:48.096731+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain GCUM on an explicitly identity-disjoint split of CSG with a fixed seed, repeated for ten runs, and compare the mean Rank-1 to UMSOT's 93.6%; if the mean falls to or below that value, the claimed generalization advantage is not demonstrated.","supporting_citations":[{"cited_title":"Group re-identification: Leveraging and integrating multi-grain information,","cited_arxiv_id":null,"evidence_quote":"Introduces the multi-grain group-reID method LIMI that this work compares against and builds on."},{"cited_title":"Learning multi-attention context graph for group-based re-identification,","cited_arxiv_id":null,"evidence_quote":"Introduces the CSG dataset and the multi-attention context graph baseline MACG."},{"cited_title":"Associating groups of people","cited_arxiv_id":null,"evidence_quote":"Introduces the iLIDS-MCTS dataset and the earliest cross-camera group association protocol."},{"cited_title":"Group re- identification via unsupervised transfer of sparse features encoding,","cited_arxiv_id":null,"evidence_quote":"PREF, a hand-crafted baseline used to show the large margin of learned text descriptions."},{"cited_title":"Uncertainty modeling with second-order transformer for group re-identification,","cited_arxiv_id":null,"evidence_quote":"The second-order-transformer uncertainty baseline SOT that the method must surpass."},{"cited_title":"Uncertainty modeling for group re-identification,","cited_arxiv_id":null,"evidence_quote":"UMSOT, the strongest single-dataset baseline; the reported gains are measured against its numbers."}],"review_version":1}