{"id":"cb191bb8-1b6c-4ae7-9979-7fa7533716fe","arxiv_id":"2504.19256","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A lightweight four-view RGBD classifier with entropy-weighted view fusion reports 95.6% on ModelNet40 and 85.1% on OmniObject3D, edging out heavier baselines.","lead":"This paper combines small convolutional encoders, vision transformers, and an entropy-weighted view fusion step into a four-view RGBD network that reports 95.6% on ModelNet40 and 85.1% on OmniObject3D. It targets robot perception, where a lightweight, fast, few-view classifier is practically useful.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim rests on unverified baseline retraining; Section IV's 'consistent viewpoint configurations' is contradicted by Table VI's 12-view vs 4-view comparison.","rationale":"The reader's weakest assumption identifies baseline fairness on OmniObject3D; I agree this is the most load-bearing issue for the headline claim. However, I would sharpen it: the paper itself claims 'consistent viewpoint configurations' (Section IV) but Table VI uses 12 views for all baselines and 4 for LM-MCVT, so the claim is internally inconsistent. This strengthens the concern that the OmniObject3D baselines may also not be equivalently configured. The GEEF fusion formula (Eq. 1) is another fixable gap, but it affects reproducibility rather than the fairness of the comparison; the central claim would still be wrong if the baselines were unfairly trained. I recommend the verdict remain CONDITIONAL pending the reproducibility test.","tokens_in":11765,"tokens_out":10279,"duration_ms":104618,"concrete_test":"Download the official MVT and SMV code. Retrain both on OmniObject3D using exactly the same four-view geometry, class-balanced 5-fold splits, optimizer settings, and training budget as LM-MCVT, running at least three seeds per method. If either baseline reaches LM-MCVT's 85.1% within one standard deviation, the claimed superiority is not robust. Separately, train LM-MCVT with 12 views on ModelNet40 and compare against VGP at 12 views to check whether the 'consistent viewpoint configurations' statement in Section IV holds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Section IV-E-1) that LM-MCVT 'sets a new benchmark' depends entirely on the comparison tables. The paper asserts in the introduction to Section IV that comparisons were made 'under consistent viewpoint configurations,' yet Table VI lists all view-based baselines at 12 views (MVCNN 80) while LM-MCVT uses 4; no same-view-count ModelNet40 comparison is reported. On OmniObject3D (Table VII), the only protocol statement is 'we trained all multi-view methods from scratch,' with no information about the number of views given to MVT/SMV, the 5-fold split assignment, whether official implementations or re-implementations were used, or the hyperparameter search budget. MVT and SMV were originally designed for 12 views; feeding them 4 views may systematically handicap their view-selection mechanisms. The reported OmniObject3D margins (85.1 vs 70.7 and 81.0) are large, but if the baselines were disadvantaged, these margins are artifacts of comparison design. The ModelNet40 margin over VGP is only 0.29 points, within typical run-to-run variance, and no standard deviation or multi-seed results are reported. Thus the SOTA claim is load-bearing on an unverified assumption of fair retraining.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LM-MCVT, a lightweight multi-view RGBD 3D object recognition architecture that combines pre- and middle-residual convolutional encoders with local and global transformers, and introduces a Globally Entropy-based Embeddings Fusion (GEEF) method for view aggregation. Experiments on ModelNet10, ModelNet40, and OmniObject3D report accuracies of 98.5%, 95.6%, and 85.1% respectively, using only four views and 10.5M parameters, together with a real-robot demonstration. The central claim is that LM-MCVT surpasses state-of-the-art view-based methods in both accuracy and efficiency.","tokens_in":11989,"tokens_out":4020,"duration_ms":40980,"significance":"If fully substantiated, the paper would make a useful practical contribution: a four-view RGBD model with 10.5M parameters and 7.0 ms inference time that outperforms much larger view-based models such as MVT and SMV on OmniObject3D, and that matches or beats 12-view baselines on ModelNet40. The component ablations and the viewpoint-structure robustness experiments are informative, and the real-robot deployment adds credibility to the efficiency claims. However, the significance currently rests on three unverified pillars: the GEEF entropy formula must be well-defined and reproducible, the OmniObject3D baseline comparisons must be shown to be fair, and the 5-fold results must be accompanied by variance information. The paper also deserves credit for reporting parameter counts and inference times, and for including ablations of the fusion strategy and architectural components.","major_comments":[{"comment":"The entropy H^cls_v(c_vj) in Eq. (1) is not computable as written. No probability distribution p is defined; the summation index j runs over views, while the argument c_vj is a single class-token vector, so the notation mixes per-view and global quantities. Since GEEF is the paper's central contribution and the weights w_vj are used in Eq. (2), the authors must define p explicitly (for example, as a softmax over the class-token vector components or over the view dimension) and correct the index structure. Without this, the fusion method cannot be reproduced or compared against alternatives.","section":"Section III-E, Eq. (1)"},{"comment":"The protocol for retraining MVT and SMV on OmniObject3D is insufficient. The statement 'we trained all multi-view methods from scratch' leaves unspecified the number of views presented to each baseline, the exact 5-fold split assignment, whether official implementations or re-implementations were used, and the hyperparameter search budget. Since MVT and SMV were originally designed for 12-view inputs, feeding them four views may systematically handicap their view-selection mechanisms. The large OmniObject3D margins (85.1 vs. 70.7 for MVT and 81.0 for SMV) therefore cannot be taken as evidence of superiority until the comparison is shown to be fair. Please report the full training and evaluation protocol, or temper the state-of-the-art claim accordingly.","section":"Section IV-E-2, Table VII"},{"comment":"The 5-fold cross-validation on OmniObject3D is reported only as single accuracy numbers, with no standard deviation, per-fold breakdown, or number of random seeds. Without this information, the reader cannot assess whether the difference between LM-MCVT (RGBD, 85.1%) and SMV (RGBD, 81.0%) is stable or within run-to-run variation. Please add per-fold results and variance estimates, or at least multiple-seed statistics, to support the robustness claim.","section":"Section IV-A and Table VII"},{"comment":"The claim of a new benchmark on ModelNet40 is not supported by a same-view-count comparison. All view-based baselines in Table VI use 12 views (MVCNN uses 80), while LM-MCVT uses four, which contradicts the paper's statement in the introduction to Section IV that comparisons were made 'under consistent viewpoint configurations.' Moreover, the 0.29-point margin over VGP on ModelNet40 is within typical run-to-run variance, and no standard deviation is reported. A fair comparison would include baselines evaluated at four views, or at least a discussion of how view count affects each baseline's accuracy.","section":"Section IV-E-1, Table VI"}],"minor_comments":[{"comment":"The merged 'Params' and 'Views' cells for the three LM-MCVT rows are easy to misread; please make explicit that Params = 10.5M and Views = 4 apply to all three input modalities.","section":"Table VI"},{"comment":"The caption contains the typo 'Rivz'; this should be 'RViz'.","section":"Fig. 6 caption"},{"comment":"The five random four-view structures are shown in figures, but the exact elevation angles and in-plane rotations are not specified; please state the generation procedure so that the robustness claim is reproducible.","section":"Section IV-D, Tables IV-V"},{"comment":"The description of how 'synthesized patch features' are formed is ambiguous because the middle-residual block excludes class tokens while the global transformer input later concatenates them; please clarify the tensor shapes and the concatenation order.","section":"Section III-D"},{"comment":"The fusion-strategy ablation is performed only on ModelNet10 and with the pre- and middle-residual encoders removed; please note this limitation explicitly, since the behavior of GEEF with the full architecture is not separately ablated.","section":"Section IV-B"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the proposed architecture may be a useful contribution if the experiments are made reproducible. The main risk is that the state-of-the-art claims depend on baseline comparisons whose fairness is not documented; I would encourage the editors to require the authors to release code and a precise description of the baseline training setup, as this is the only way to verify the central claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent engineering paper with a nice robot demo, but the headline claim that LM-MCVT \"sets a new benchmark\" is not supported by the evidence as presented. The architecture is a mix of known ingredients — residual convs around a ViT, local/global transformer pairs, class-token fusion — and the main novelty, GEEF, is entropy-weighted fusion over class tokens, which is a standard idea with an underspecified formula. What the paper does well: the ablation study is thorough (fusion variants, component removal, view-count and view-structure robustness), the four-view RGBD setting is practically relevant, and the real-robot pick-and-place test is a genuine attempt at validation. The efficiency numbers (10.5M params, 7 ms) are attractive.\n\nThe soft spots are real and load-bearing for the SOTA claim. First, Eq. (1) defines H as −Σ p log p but never says what p is. The class token is an embedding, not a distribution; until p is defined the equation is incomplete. Second, the paper says comparisons were made under \"consistent viewpoint configurations,\" but Table VI lists all published baselines at 12 views while LM-MCVT uses 4. If those numbers are copied from the original papers, that is fine as a reference table, but it is not a same-protocol comparison. Third, on OmniObject3D the only protocol statement is \"we trained all multi-view methods from scratch\" — no view count, no splits, no hyperparameters, no mention of official code. Since MVT and SMV were designed around 12 views, feeding them 4 could handicap them. The 4–15 point margins on OmniObject3D are suspicious precisely because the protocol is invisible. Fourth, the ModelNet40 margin over VGP is 0.29 points, which is within run-to-run noise; no standard deviation is reported even for the 5-fold OmniObject3D result.\n\nNone of this makes the method useless. If the authors disclose the full retraining setup and release code or checkpoints, the empirical claims can be checked. As it stands, the paper is a solid engineering contribution with unreliable comparison claims. I would send it to peer review but with a demand for a complete protocol description, same-view-count comparisons (or an explicit statement that baselines use published 12-view numbers), a definition of p in GEEF, and variance estimates. The robot demo is a plus but does not substitute for a sound comparison.\n\nWho this is for: people building lightweight RGBD multi-view classifiers for robotics. They will get value from the architecture and ablations even if the SOTA claim is discounted. I would not cite the headline numbers until the protocol issue is resolved, but I would cite the architecture/ablation parts in a related work section.\n\nRecommendation: conditional accept with major revisions; the core method is reasonable but the evaluation needs to be made honest and reproducible.","headline":"Useful lightweight RGBD multi-view classifier with careful ablations, but the state-of-the-art claim is undercut by inconsistent view counts and an undocumented baseline-retraining protocol.","tokens_in":12583,"tokens_out":3416,"would_cite":false,"duration_ms":31224,"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":"A lightweight network using four RGB-D views can match or beat view-based 3D classifiers that use twelve or more views, the paper reports on ModelNet and OmniObject3D.","keywords":["3D object recognition","multi-view learning","RGB-D fusion","vision transformer","entropy-based fusion","lightweight model","robotic perception","ModelNet40"],"falsifier":"Retrain MVT and SMV on the same four-view RGB-D renderings and the same ModelNet and OmniObject3D splits, with the same hyperparameter budget, and measure accuracy; if either baseline reaches or exceeds 95.6% on ModelNet40 or 85.1% on OmniObject3D under that protocol, the claimed state-of-the-art result would not hold.","tokens_in":11459,"feed_emoji":"🤖","tokens_out":7398,"duration_ms":67363,"temperature":0.7,"pith_summary":"The paper claims that a lightweight network can recognize 3D objects from only four RGB-D views as accurately as heavier view-based models that use twelve or more views. The proposed LM-MCVT stacks convolutional encoders before and between transformer stages so that local details survive global attention, and it introduces Globally Entropy-based Embeddings Fusion (GEEF), which weights each view by the entropy of its class token before merging RGB and depth features. Across ModelNet10, ModelNet40, and the real-scanned OmniObject3D dataset, the reported accuracies are 98.5%, 95.6%, and 85.1%, with 10.5 million parameters and about 7 milliseconds per RGB-D instance. A dual-arm robot using the fine-tuned model succeeded in ten consecutive pick-and-place trials, which the paper offers as evidence of practical deployability.","feed_headline":"Four RGB-D views top heavier 3D object classifiers","feed_subtitle":"Lightweight network hits 95.6% on ModelNet40 and 85.1% on real-scanned OmniObject3D with entropy-weighted view fusion.","key_machinery":"The load-bearing mechanism is GEEF (Globally Entropy-based Embeddings Fusion), which turns each view's class-token entropy into a normalized weight and fuses class tokens by that weight while average-pooling the remaining patch embeddings; RGB and depth streams are fused separately and concatenated. The rest of the architecture exists to feed GEEF good tokens: pre-residual convolutional encoders and middle-residual convolutional encoders restore local detail, local transformers encode patch relations within a view, and global transformers let views exchange information through class tokens. Together they give the model a lightweight profile of 10.5 million parameters.","core_discovery":"The central claim is that coordinated use of convolutional feature extraction, transformer-based global reasoning, and entropy-weighted view fusion produces a 3D object recognizer that is both more accurate and cheaper than the published comparison methods. With four rendered RGB-D views, LM-MCVT reaches 98.5% on ModelNet10, 95.6% on ModelNet40, and 85.1% on OmniObject3D in five-fold cross-validation, while using 10.5 million parameters and 7.0 ms per instance. The paper attributes the improvement over average-pooling fusion to GEEF, which normalizes the entropy of each view's class token into a weight and uses it to combine the views, so views carrying sharper or more distinct information dominate the fused representation.","pith_inferences":["If the four-view protocol is taken at face value, the implication extends beyond classification: dense multi-view capture may be unnecessary whenever depth is available, so latency budgets in robotics can be spent on other perception stages.","GEEF is a generic fusion operator: it could be applied wherever multiple views or frames produce class tokens, such as video action recognition, multi-camera detection, or active perception, where viewpoint informativeness varies.","A natural test of the paper's mechanistic claim is to visualize the learned entropy weights per view; if high-entropy views are the ambiguous ones, entropy weighting is doing uncertainty-aware pooling, whereas if high-entropy views are the informative ones, it is doing saliency weighting.","The absence of a described baseline protocol means the strongest testable extension is an independent reproduction with matched views, splits, and tuning, which would separate the contribution of the architecture from the contribution of the evaluation setup."],"forward_implications":["Four views are close to the practical ceiling: on ModelNet10, RGB-D accuracy moves from 98.5% at four views to 98.9% at twelve views while per-instance time grows from 7.0 ms to 29.7 ms.","Depth consistently helps: on ModelNet10, adding depth raises accuracy at every view count, from 93.2% to 95.5% in the single-view case.","The entropy-weighted fusion beats the alternatives it was tested against: GEEF reaches 96.6% on ModelNet10 RGB-D versus 95.4% for the best averaging-based embedding fusion.","The same four-view model transfers to real-scanned data, where it reports 85.1% accuracy under five-fold cross-validation, and a fine-tuned version reached 99.3% on synthetic household objects and succeeded in ten real pick-and-place trials.","At 10.5 million parameters and 7.0 ms per RGB-D instance, the model sits in a practical range for real-time robotic perception."],"supporting_citations":[{"why":"Defines the MVT baseline and the multi-view transformer structure that LM-MCVT extends with convolutional encoders; also supplies the ModelNet view settings and train-test splits used in the main comparison.","marker":"[14]"},{"why":"SMV is the strongest view-based comparator in the OmniObject3D table and a published method the paper claims to surpass.","marker":"[12]"},{"why":"Introduces the ModelNet10 and ModelNet40 datasets and the accuracy measures used for the headline results.","marker":"[22]"},{"why":"Provides the real-scanned OmniObject3D dataset on which the five-fold cross-validation comparison is run.","marker":"[23]"},{"why":"Supplies the observation that deep vision transformers lose local feature detail, the premise behind the pre- and middle-residual convolutional encoders.","marker":"[18]"},{"why":"MVCNN is the dense multi-view baseline (80 views) that LM-MCVT outperforms with four views, supporting the efficiency claim.","marker":"[27]"},{"why":"Establishes class-token embeddings in vision transformers, the representation that GEEF entropy-weights and fuses.","marker":"[24]"}],"fun_headline_variants":["Entropy-weighted views lift 3D recognition to 95.6%","Lightweight CNN-transformer hits 95.6% on ModelNet40","Four-view entropy fusion tops 3D classifiers","Multi-view model: 95.6% accuracy, 7 ms per object","GEEF fusion: sharper views dominate 3D recognition"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The accuracy comparison rests on the assumption that MVT and SMV were retrained under exactly the same four-view rendering, the same five-fold splits, and a comparable tuning budget; the paper does not describe that protocol.","fun_headline_variants_meta":{"raw":{"variants":["Entropy-weighted views lift 3D recognition to 95.6%","Lightweight CNN-transformer hits 95.6% on ModelNet40","Four-view entropy fusion tops 3D classifiers","Multi-view model: 95.6% accuracy, 7 ms per object","GEEF fusion: sharper views dominate 3D recognition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000174,"raw_usage":{"total_tokens":1268,"prompt_tokens":919,"completion_tokens":349,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":535,"completion_tokens_details":{"reasoning_tokens":256}},"tokens_in":535,"tokens_out":349,"duration_ms":3437,"temperature":1.0,"reasoning_tokens":256,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:58:04.932542+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain MVT and SMV on the same four-view RGB-D renderings and the same ModelNet and OmniObject3D splits, with the same hyperparameter budget, and measure accuracy; if either baseline reaches or exceeds 95.6% on ModelNet40 or 85.1% on OmniObject3D under that protocol, the claimed state-of-the-art result would not hold.","supporting_citations":[{"cited_title":"MVT: multi-view vision transformer for 3d object recognition,","cited_arxiv_id":null,"evidence_quote":"Defines the MVT baseline and the multi-view transformer structure that LM-MCVT extends with convolutional encoders; also supplies the ModelNet view settings and train-test splits used in the main comparison."},{"cited_title":"Selective multi- view deep model for 3d object classification,","cited_arxiv_id":null,"evidence_quote":"SMV is the strongest view-based comparator in the OmniObject3D table and a published method the paper claims to surpass."},{"cited_title":"Do vision transformers see like convolutional neural networks?","cited_arxiv_id":null,"evidence_quote":"Supplies the observation that deep vision transformers lose local feature detail, the premise behind the pre- and middle-residual convolutional encoders."}],"review_version":1}