{"id":"6007d15a-b975-4671-ad2b-6eebf5272097","arxiv_id":"2411.10077","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"HMDMV trains a multi-view classifier on all view subsets with uncertainty-weighted fusion and hierarchical mutual distillation, improving accuracy and supporting flexible view counts at inference.","lead":"This paper proposes HMDMV, a training method that makes a multi-view classifier learn from every possible subset of the available camera views, then aligns those subset predictions through mutual distillation. A smart generalist might read it to see a practical recipe for handling real-world images where the number of available angles changes from sample to sample, such as in medical or 3D recognition data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim confounded by training compute: HMDMV enumerates all 2^n-1 view subsets (Eqs. 2-4), ~2-4x MV-HFMD's transformer FLOPs at n=3-4, without FLOP/epoch reporting or a compute-matched baseline; the advertised light version is never evaluated.","rationale":"The reader's weakest assumption correctly identifies the most load-bearing threat to the central claim. The paper's Tables 1-2 are the sole support for SOTA accuracy, and the comparison is not controlled for training compute. HMDMV's design explicitly computes predictions for all C(n,k) subsets at every k (Eq. 2, 4), and the loss in Eq. 9 supervises every subset, so each training step consumes far more transformer computation than MV-HFMD, which only processes single views and the full view set. The reported gains grow with n (1.46%, 3.56%, 3.38% on Hotels-8k; 0.85%, 2.31%, 3.80% in Table 2), which is consistent with an exponentially growing compute advantage. The ablation study Table 3 demonstrates that each proposed component helps, but it never tests HMDMV against MV-HFMD at equal FLOPs or epochs. The abstract's promised light version would be precisely the experiment needed to decouple algorithm from compute, and its absence is a self-admitted gap rather than an artifact of the review pipeline. Without that experiment, a reasonable alternative explanation for the SOTA numbers exists. The concern is not that the method is wrong, but that the paper's strongest claim is not yet established; the appropriate disposition is the reader's CONDITIONAL, so no verdict change is needed.","tokens_in":14190,"tokens_out":7276,"duration_ms":71131,"concrete_test":"Run the Hotels-8k and GLDv2 3-view/4-view comparisons with the abstract's \"light version\" (randomly sample one k-subset per step) while budgeting training FLOPs equal to MV-HFMD (e.g., match summed S^2-weighted transformer forwards per sample); run 3 seeds and report mean +/- std. If HMDMV-light fails to reproduce the Table 1-2 margins over MV-HFMD, the SOTA claim is compute-driven rather than due to the proposed distillation/weighting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central \"state-of-the-art\" claim (Tables 1-2) rests on comparisons to MV-HFMD and others, but HMDMV trains on every subset in C_k for k=1..n (Eq. 2-4), i.e., 2^n-1 transformer forward passes per sample. At n=4 this is 15 forwards vs the 5 (4 single + 1 full) used by MV-HFMD; weighting by sequence length (sum_k C(n,k) k^2 vs n + n^2) makes HMDMV roughly 4x more transformer FLOPs at n=4 and about 2x at n=3. The paper reports no training epochs, FLOPs, or seeds. The ablation (Table 3) only removes components within HMDMV and cannot distinguish the gap to MV-HFMD from extra compute. The abstract promises a light version with random subset sampling, but no such experiment appears anywhere in the paper, so there is no evidence the claim survives at matched compute. In addition, several winning margins are <=0.15 pp (Carvana 2-view top-1, GLDv2 4-view top-5) with no error bars, so the SOTA claim is not statistically established. The compute confound is therefore the load-bearing unaddressed assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HMDMV, a multi-view classification method that trains a hybrid CNN-transformer to produce predictions for every non-empty subset of n input views. Within each k-view combination set, predictions are fused by inverse-uncertainty weighting based on entropy plus cross-entropy, and hierarchical mutual distillation aligns single-view and partial multi-view predictions with the full multi-view prediction through a KL-based loss with a per-scale weight. Experiments on Hotels-8k, GLDv2, and Carvana compare fixed-view and variable-view inference against MVCNN, GVCNN, MVT, and MV-HFMD, with ablations and inference-time measurements.","tokens_in":14515,"tokens_out":5876,"duration_ms":60392,"significance":"If the claims are established, HMDMV extends mutual-distillation multi-view fusion from single-view/full-view pairs to all view subsets, and its flexible-inference property (handling fewer or more views than seen in training) is practically useful for unstructured data. The method is clearly specified, the code is released, and the evaluation spans both unstructured and structured settings with an ablation study. However, the central state-of-the-art claim is not yet demonstrated because training compute is neither matched nor reported, the inference-time weighting rule is underspecified in a way that depends on ground-truth labels in Eq. (5), and no statistical variability is reported for the main tables. The flexible-inference contribution is credible and independent of the SOTA accuracy debate.","major_comments":[{"comment":"The reported accuracy gains over MV-HFMD are confounded by a training-compute gap. HMDMV evaluates all 2^n-1 subsets C_k per training sample, whereas MV-HFMD evaluates n single views and the full n-view combination; given the sequence lengths in Eq. (3), HMDMV's transformer FLOPs are approximately twice those of MV-HFMD at n=3 and about four times at n=4. The paper reports no training epochs, FLOPs, or wall-clock training time, and the authors themselves note the increased training cost in Section 7. The abstract additionally promises a light version with random subset sampling, but no such experiment appears anywhere in the paper. To support the state-of-the-art claim, the authors need a compute-matched comparison (e.g., MV-HFMD trained for as many FLOPs), a cost-versus-accuracy analysis, or an evaluated light version; otherwise the gains in Tables 1-2 cannot be attributed to the proposed components.","section":"Section 3.1, Eqs. (2)-(4); Tables 1-2"},{"comment":"The uncertainty weight w_i^k is defined through u_i^k in Eq. (5), which includes the cross-entropy term q_j log(p_i^k(j)+epsilon) with q_j the ground-truth one-hot label. The same weights are then used in Eq. (7) to compute P(C_k), the fusion mechanism used for the results in Tables 1-2 and for the unstructured-view ensembling described in Section 5. No procedure is given for computing these weights at inference time, when q_j is unavailable. The authors must specify the test-time weighting rule (e.g., entropy-only, equal weights, or a different unsupervised estimate) and, if it differs from Eq. (5), evaluate that rule explicitly; otherwise the reported fusion results are not reproducible.","section":"Section 3.2, Eqs. (5)-(7); Section 5"},{"comment":"The state-of-the-art claim is not statistically supported because no error bars, multiple-seed results, or significance tests are reported. Several margins are small: Carvana 2-view top-1 is 93.33% vs 93.23% for MV-HFMD, GLDv2 4-view top-5 is 99.95% vs 99.89%, and on GLDv2 2-view top-1 HMDMV is below MV-HFMD (96.35% vs 96.48%). Without repeated runs, these differences are within plausible run-to-run variability. Please add mean and standard deviation over at least three seeds and a significance test for the headline comparisons.","section":"Tables 1-2"}],"minor_comments":[{"comment":"The checkmark columns are not aligned with the component names in the typeset table, and two rows show three checkmarks that appear to represent different component sets; please reformat the table so each row explicitly names the enabled components.","section":"Table 3"},{"comment":"There are apparent typographical errors in the inference-time numbers, such as '430 .27', '530 .99', '648 .06', and '1684 .37', with stray spaces and missing digits; also, the column headers mix 'nt > n' and 'nt ≥ n'.","section":"Table 5"},{"comment":"The text calls λ, τ, and α three primary hyperparameters, but τ is fixed at 4 and λ_k is a scaled version of λ; the term 'adaptive λ' is confusing because the scale is set to a fixed α=1.2 and λ=0.1 after tuning rather than adapted during training.","section":"Section 4.1, Eq. (12)"},{"comment":"The abstract promises a light version with reduced training cost by random subset sampling, but the experiments and Section 7 only discuss future work on combination optimization; please either add the promised light-version experiments or remove the claim from the abstract.","section":"Abstract and Section 7"}],"recommendation":"major_revision","confidential_remarks":"The core idea is sound and within scope for a computer vision journal; the flexible-inference property is a genuine strength. The main risks are the training-compute confound and the underspecified inference-time weighting in Eq. (5). Both are addressable with additional experiments and clarifications, so I do not recommend rejection. Please ask the authors to report training FLOPs/epochs, add a compute-matched or evaluated light-version comparison, and clarify how uncertainty weights are computed when labels are not available."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: HMDMV is a well-specified, plausible method that deserves a real review, but the SOTA claim is not yet established. The new thing is training on all possible view subsets with uncertainty-weighted fusion and hierarchical mutual distillation from each subset level to the full set. That is a clean extension of MV-HFMD, and the paper does it carefully: equations are explicit, the architecture is described, and experiments cover structured and unstructured data, with a useful flexible-inference protocol. I also credit the inference-time table, which gives actual measurements rather than hand-waving.\n\nThe soft spots are the ones in the stress test and they are real. HMDMV enumerates 2^n - 1 subsets per sample during training. At n=3-4 that is roughly 2-4x the transformer forward passes of MV-HFMD per step, and the paper reports no epochs, FLOPs, or compute-matched baseline. The ablation in Table 3 only removes components within HMDMV, so it cannot separate the method from the extra compute. The wins over MV-HFMD are often small (0.1-0.15 pp, e.g. Carvana 2-view top-1 and GLDv2 4-view top-5) and there are no error bars or seeds anywhere, so the SOTA claim is not statistically grounded. The abstract also promises a light version with random subset sampling, but that experiment is not in the paper; either it needs to be reported or removed. These are fixable in revision, but they are load-bearing for the main claim.\n\nThe math and loss design are standard and internally consistent. The uncertainty weighting is supervised via the ground-truth term in Eq. 5, so it is not circular; it is just a training mechanism. The citation pattern looks fair, with the main comparison to MV-HFMD clearly drawn.\n\nWho is this for? Anyone working on multi-view classification with variable view counts, especially medical imaging or 3D recognition. It is a useful recipe even if the SOTA gap narrows under matched compute. I would send it to review, with the request that the authors add seeds/error bars, report training cost, add a compute-matched baseline or at least a random-subset training variant, and address the light version. If they do that, the claim will be in much better shape.","headline":"Solid training recipe for variable-view classification, but the SOTA claim is currently confounded by training compute and missing error bars.","tokens_in":15016,"tokens_out":1486,"would_cite":false,"duration_ms":15682,"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":"Training on all possible view subsets lifts multi-view classification accuracy and enables flexible inference with any number of test views.","keywords":["multi-view learning","hierarchical mutual distillation","uncertainty weighting","view combinations","multi-view fusion","flexible inference","image classification","knowledge distillation"],"falsifier":"Train the strongest baseline with the same number of view-combination forward passes per optimization step as HMDMV (or match total FLOPs), then compare top-1 accuracy on Hotels-8k; if the gap disappears, the method's benefit over a similarly expensive ensemble of all view subsets is unestablished.","tokens_in":14008,"feed_emoji":"🖼️","tokens_out":3961,"duration_ms":37393,"temperature":0.7,"pith_summary":"This paper proposes HMDMV, a multi-view image classification method that trains on all possible subsets of the input views—single view, every partial combination, and the full set—rather than only on individual views and the complete set. For each combination set it fuses the member predictions with uncertainty weights, then applies bidirectional mutual distillation between each combination level and the full-multi-view prediction. Across structured (Carvana) and unstructured (Hotels-8k, GLDv2) datasets, the method reports top-1 accuracy gains over prior fusion and distillation baselines, with the largest gains when three or four views are available. The practical payoff is flexibility: after training with n views, the model can classify with any number of test views, fewer or greater, without retraining.","feed_headline":"All view subsets sharpen multi-view classification","feed_subtitle":"HMDMV distills across single, partial, and full view sets, and accepts any number of views at test time.","key_machinery":"The load-bearing object is the combination set C_k—the collection of all k-view subsets of the n input views—together with the uncertainty-weighted fusion P(C_k) and the hierarchical mutual distillation loss L_hmd, which aligns each single/partial combined prediction with the full multi-view prediction by bidirectional KL divergence. This machinery propagates supervision through the exponential family of view subsets so the network learns consistent representations across all granularities of view availability.","core_discovery":"The central claim is that inter-view relationships are better learned by explicitly supervising every view-combination level than by distilling only between single-view and full-multi-view predictions. HMDMV builds a hybrid CNN-Transformer that outputs a prediction for each k-view subset; within each level the predictions are averaged with weights inversely proportional to an entropy-plus-cross-entropy uncertainty. These combined level predictions are then mutually distilled against the full multi-view prediction in both directions, with the distillation weight scaled by k^α. The paper reports that this scheme achieves the highest top-1 and top-5 accuracies on Hotels-8k, GLDv2, and Carvana across 2-, 3-, and 4-view settings, and that it degrades gracefully when individual test samples have more or fewer views than the training configuration.","pith_inferences":["If the reported gains are causal, the same combination-set distillation could transfer to multi-modal inputs (e.g., RGB plus depth plus text), where the partial-combination levels correspond to modality subsets.","The strongest baseline, MV-HFMD, is trained with fewer forward passes per step, so an equal-compute comparison would clarify how much of the gain comes from the distillation and how much from ensembling many view subsets; the paper does not provide that comparison.","The uncertainty weighting uses ground-truth cross-entropy during training, meaning the weights are not purely uncertainty-based; at test time the fusion must fall back to entropy alone, so a test-time-only evaluation could quantify how much of the benefit relies on label access.","A testable extension: vary the number of training views and measure whether the marginal accuracy gain from HMDMV saturates; the method's exponential training cost makes a sampling schedule an obvious practical variant."],"forward_implications":["Inference view count becomes a free parameter: at test time the model accepts any subset of views without padding or duplication when nt ≤ n, and ensembles valid subsets when nt > n.","Performance improves as the number of views grows, and on unstructured datasets HMDMV holds accuracy where baselines decline under variable view counts.","The method is not tied to a fixed backbone, so the combination-set training and distillation losses can be ported to stronger or lighter CNN-Transformer architectures.","A light version that randomly samples view subsets per iteration reduces training cost, making the approach more practical for large numbers of views."],"supporting_citations":[{"why":"Supplies the hybrid fusion and mutual distillation baseline that HMDMV extends and compares against on all three datasets.","marker":"[3]"},{"why":"Provides the temperature-scaled KL distillation loss that the hierarchical mutual distillation term builds on.","marker":"[25]"},{"why":"Supplies the mutual/bidirectional distillation principle used to align view-combination predictions.","marker":"[64]"},{"why":"Provides the unstructured hotel-room multi-view dataset used for the main accuracy and inference-time evaluations.","marker":"[29]"},{"why":"Provides the large-scale unstructured landmark dataset used for evaluation.","marker":"[56]"},{"why":"Provides the structured 16-view vehicle dataset used for fixed-view evaluation.","marker":"[46]"},{"why":"One of the transformer-based baselines that HMDMV must beat in fixed and unstructured evaluations.","marker":"[7]"},{"why":"Classic multi-view baseline that defines the early single-architecture fusion approach compared throughout.","marker":"[50]"}],"fun_headline_variants":["Distilling every view subset sharpens multi-view learning","Train on all view combos, test on any number of views","HMDMV distills all view subsets for robust multi-view fusion","Learning from every possible view combination improves accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported accuracy gains are caused by the proposed hierarchical mutual distillation and uncertainty weighting rather than by the much larger training compute used to evaluate all 2^n - 1 view subsets per sample; the paper does not equalize FLOPs or wall-clock time across methods.","fun_headline_variants_meta":{"raw":{"variants":["Distilling every view subset sharpens multi-view learning","Train on all view combos, test on any number of views","HMDMV distills all view subsets for robust multi-view fusion","Learning from every possible view combination improves accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000406,"raw_usage":{"total_tokens":2106,"prompt_tokens":935,"completion_tokens":1171,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":1104}},"tokens_in":551,"tokens_out":1171,"duration_ms":9980,"temperature":1.0,"reasoning_tokens":1104,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:59:38.291438+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the strongest baseline with the same number of view-combination forward passes per optimization step as HMDMV (or match total FLOPs), then compare top-1 accuracy on Hotels-8k; if the gap disappears, the method's benefit over a similarly expensive ensemble of all view subsets is unestablished.","supporting_citations":[{"cited_title":"Multi-view classifica- tion using hybrid fusion and mutual distillation","cited_arxiv_id":null,"evidence_quote":"Supplies the hybrid fusion and mutual distillation baseline that HMDMV extends and compares against on all three datasets."},{"cited_title":"Deep mutual learning","cited_arxiv_id":null,"evidence_quote":"Supplies the mutual/bidirectional distillation principle used to align view-combination predictions."},{"cited_title":"The 2021 Hotel-ID to Combat Human Trafficking Competition Dataset","cited_arxiv_id":"2106.05746","evidence_quote":"Provides the unstructured hotel-room multi-view dataset used for the main accuracy and inference-time evaluations."},{"cited_title":"Google landmarks dataset v2-a large-scale benchmark for instance-level recognition and retrieval","cited_arxiv_id":null,"evidence_quote":"Provides the large-scale unstructured landmark dataset used for evaluation."},{"cited_title":"Carvana image masking challenge","cited_arxiv_id":null,"evidence_quote":"Provides the structured 16-view vehicle dataset used for fixed-view evaluation."},{"cited_title":"Multi-view convolutional neural networks for 3d shape recognition","cited_arxiv_id":null,"evidence_quote":"Classic multi-view baseline that defines the early single-architecture fusion approach compared throughout."}],"review_version":1}