{"id":"5a106bf4-54ed-45a5-8350-190cd4d71444","arxiv_id":"2506.12401","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A CNN-ViT hybrid with frequency-spatial adapters and dynamic fusion is reported to reach new high Recall@1 scores on several VPR benchmarks.","lead":"This paper proposes LGCN, a visual place recognition network that pairs a frozen vision Transformer with a CNN and adds frequency-spatial adapters plus a dynamic fusion gate. It reports top results on several VPR benchmarks, but a central fusion equation appears to ignore the CNN branch entirely.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (5) in §3.2 defines DFM as recombining only F_ViT, so the claimed CNN–ViT complementation cannot occur as written; the central mechanism is either a typo or absent, and no code or corrected equations are supplied.","rationale":"The reader's strongest claim is that LGCN's parallel CNN-ViT streams and DFM are what produce the reported state-of-the-art results. I agree that is the central assertion. I therefore looked for the point at which that assertion is least secure and found it inside the method itself: Eq. (5) in Sec. 3.2. The three-line derivation (Eqs. 3–5) is the complete mathematical specification of DFM. Eq. (3) sums the branches, Eq. (4) produces an attention map, and Eq. (5) applies that map to F_ViT twice (once with ω, once with 1−ω). The ResNet feature F'_Res appears only in Eq. (3) as an input to the gate. Thus, if the paper is taken literally, the CNN branch is not part of the fused descriptor. That makes the paper's title, abstract, and ablation interpretation unsupported. A reader could charitably call this a typo, but no corrected equation or code is provided, and the untracked learnable parameters α1 and α2 deepen the ambiguity. I also noted smaller consistency issues: the Full Model row in Table 4 reports Pitts30k R@1=95.5 while Table 2 reports Ours=95.0 for the same setting, and the text's 'significantly outperforming' is contradicted by Table 2 on MSLS val (BoQ 91.4 vs. 91.0) and by Table 3 on Eynsham (BoQ 91.5 vs. 90.7). None of these secondary issues is as damaging as Eq. (5), but they reinforce that the evaluation is not self-consistent. The reader's protocol-comparison worry is legitimate but secondary: even with perfectly matched baselines, the paper would still not explain how CNN and ViT features are actually recombined. For this reason I recommend no change to the reader's REJECT verdict.","tokens_in":12799,"tokens_out":3795,"duration_ms":42677,"concrete_test":"Implement DFM exactly as printed in Eq. (5) on Pitts30k with the same GSV-Cities training protocol, verifying whether gradients from the final descriptor reach the ResNet branch. If Eq. (5) is used verbatim, the CNN branch receives no gradient through F_fused1/F_fused2, and the '+DFM' row of Table 4 should collapse toward the ViT-only baseline. Then implement the natural correction F_fused1 = ω⊙F_ViT + (1−ω)⊙F'_Res and F_fused2 = (1−ω)⊙F_ViT + ω⊙F'_Res and rerun the same ablation. Comparing the two Recall@1 numbers settles whether the published equation is the source of the claimed fusion gain or a typo.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that LGCN's parallel CNN-ViT streams plus DFM produce state-of-the-art VPR accuracy. Section 3.2's DFM equations make this claim unverifiable. Eq. (3) forms F = F_ViT + F'_Res, then Eq. (4) computes gate ω from F. Eq. (5) then defines F_fused1 = ω⊙F_ViT and F_fused2 = (1−ω)⊙F_ViT. Both fused outputs are functions only of the ViT branch; F'_Res enters only through the gate ω. Consequently, as written, the CNN stream cannot contribute any feature content to the final descriptor beyond influencing a scalar/spatial gate on ViT features. The text explicitly claims 'enhancing the contribution of CNN features in regions rich in texture' (Sec. 3.2), which Eq. (5) cannot do. The learnable α1 and α2 introduced in the same section are never used in any equation. This is not a style issue: it is the core mechanism named in the title and abstract. The ablation row '+DFM' (Table 4) cannot validate a fusion that the equations do not describe, and the reported gains over '+CNN Stream' are therefore unexplained. Without code, the reader cannot tell whether Eq. (5) is a typographical omission of F'_Res or the actual implementation; either way the paper's central claim lacks a correct, checkable specification.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes LGCN, a visual place recognition network combining a frozen DINOv2 ViT-Base branch and a frozen DINO-pretrained ResNet-50 branch. A dynamic feature fusion module (DFM) is introduced to recombine the two branches through a gating mechanism, and frequency-spatial adapters (FSA) are inserted into the frozen ViT to add local inductive bias. A cross-image correlation-aware module borrowed from CricaVPR is used on top of the fused features to produce the final descriptor. The method is evaluated on seven VPR benchmarks (Pitts30k, Pitts250k, SPED, MSLS val, Nordland, Eynsham, St-Lucia) and reported to achieve state-of-the-art R@1 on most of them. The central claim is that the CNN-ViT feature complementation through DFM is responsible for the accuracy gains.","tokens_in":13147,"tokens_out":8730,"duration_ms":93566,"significance":"If the claimed mechanism were realized as described, the paper would address a real and active problem: combining local CNN features and global ViT features for visual place recognition. The use of parameter-efficient adapters for a frozen backbone is also timely. The evaluation is broad and follows the common GSV-Cities training protocol, so the empirical comparison is not circular. However, the significance is substantially undermined because the core fusion equations, as written, do not implement the advertised CNN-ViT complementation, and no code or corrected specification is supplied. The contribution relative to CricaVPR is also not clearly demarcated, since the cross-image module is adopted from that work.","major_comments":[{"comment":"The DFM as specified does not fuse CNN features into the output. Eq. (3) forms F = F_ViT + F'_Res and Eq. (4) computes the gate ω from F, but Eq. (5) defines Ffused1 = ω⊙F_ViT and Ffused2 = (1−ω)⊙F_ViT. Both fused outputs are functions only of F_ViT; F'_Res appears only inside ω, where it can modulate a gate but cannot contribute any feature content. The text at the same point claims that the mechanism 'enhanc[es] the contribution of CNN features in regions rich in texture,' which Eq. (5) cannot do. The learnable parameters α1 and α2 are introduced in the same section but are never used in any equation. Because the DFM is the central contribution named in the title, abstract, and ablation study, this is not a cosmetic typo: the paper does not provide a correct, checkable specification of its core mechanism, and the +DFM row in Table 4 cannot validate a fusion that the equations do not describe.","section":"Section 3.2, Eqs. (3)-(5)"},{"comment":"The reported R@1 for the full model on Pitts30k is inconsistent across tables: Table 2 lists 95.0 for 'Ours' while Table 4 lists 95.5 for the 'Full Model (Ours)'. The paper also claims in the abstract that LGCN 'consistently outperforms' prior methods, but Table 2 shows BoQ achieving a higher R@1 on MSLS val (91.4 vs 91.0) and Table 3 shows BoQ tied on St-Lucia (99.9). These discrepancies and exceptions need to be stated precisely and reconciled.","section":"Section 4.3 / Section 4.4, Tables 2 and 4"},{"comment":"The claim of 'significantly outperforming' is not supported by the reported margins. Against CricaVPR on Pitts30k the improvement is 0.1 percentage point and against SelaVPR on Pitts250k it is 0.3 point. No standard deviations, multiple-run statistics, or significance tests are reported, so these differences may be within run-to-run variation. Please either soften the wording or provide statistical evidence.","section":"Section 4.3, Table 2"},{"comment":"The architecture description is internally inconsistent about the ViT patch size. Section 3.1 states that input images are divided into 16×16 patches and that F_ViT has shape B×16×16×768, while Section 4.2 says the backbone is ViT-B/14. A 224×224 input with patch size 14 yields 16×16 tokens, so the shape is consistent with a patch size of 14, not 16. Please clarify which ViT variant is used and make the patch-size statements consistent, since this affects reproducibility.","section":"Section 3.1 / Section 4.2"}],"minor_comments":[{"comment":"The notation switches between w1/w2 in the equation and ω1/ω2 in the text; please use consistent symbols.","section":"Section 3.2, Eq. (4)"},{"comment":"The sentence 'BoQ [43] designed learnable global query vectors' cites reference [43], which in the same paragraph is CricaVPR; the correct citation is [63]. This also affects the related-work narrative.","section":"Section 2.2"},{"comment":"The cross-image correlation-aware module is adopted from CricaVPR, but the paper does not clearly state in the method section that this is a borrowed component rather than a new contribution; please mark the provenance in Fig. 4 and its surrounding text.","section":"Section 4.2 / Fig. 4"},{"comment":"Training details are incomplete: no loss function, number of epochs, or evaluation protocol for the additional datasets (Nordland, Eynsham, St-Lucia) are given, so the reader cannot reproduce the results.","section":"Section 4.2"},{"comment":"The bar chart in Fig. 2 is not referenced in the text, and its caption lists SPED but not the other datasets; please add a reference and complete the caption.","section":"Fig. 2"},{"comment":"The caption mentions 'SFRS' as a compared method, but SFRS is not described or listed in the baselines.","section":"Fig. 6"},{"comment":"The frequency-spatial adapter is described only in words; no equations or parameter counts are given, making it difficult to assess the claimed 'lightweight' property.","section":"Section 3.3"}],"recommendation":"reject","confidential_remarks":"The manuscript does not clearly separate what is new from the CricaVPR cross-image module, and no code or corrected equations are provided to verify the central fusion mechanism. If the authors can supply a corrected specification and code, a resubmission might be considered, but as submitted the core claim is not checkable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper reports strong benchmark numbers, but the DFM equations in Sec. 3.2 don't do what the text claims, and the paper contradicts its own tables. If Eq. (5) is a typo, it's fixable; if it's the actual implementation, the method isn't what the title says.\n\nWhat's actually new: the specific combination of DINOv2 ViT, DINO ResNet-50, frequency-spatial adapters, and a gated fusion module, evaluated on seven VPR benchmarks. That combination is not in the prior literature. The ablation study is a genuine attempt to isolate contributions, and the gains over CricaVPR are typical for the field (0.1–2 points). The frequency-spatial adapter is a reasonable transfer of ideas from refs 52–54 into VPR, and the cross-image encoder is borrowed cleanly from CricaVPR.\n\nSoft spots, in order of seriousness. Eq. (5) computes both fused branches from F_ViT alone; the CNN stream enters only through the gate ω. That means the final descriptor has no direct CNN feature content, contradicting the claimed local-global complementation. The learnable α1 and α2 are introduced but never used in any equation. The paper says 'consistently outperforms existing approaches' but its own Table 2 shows BoQ ahead on MSLS val R@1 (91.4 vs 91.0), and Table 3 shows BoQ ahead on Eynsham R@1 (91.5 vs 90.7). The implementation details say ViT-B/14, but Sec. 3.1 says 16×16 patches. No code is provided, so a reviewer cannot tell which description matches the trained model.\n\nTo be fair, the empirical work is not obviously suspect: the numbers sit in a plausible range and the ablations have internal consistency. Low circularity burden — trained on GSV-Cities, tested on public benchmarks. The core flaw is a specification problem, not necessarily a data problem.\n\nFor whom: someone working on VPR feature fusion might find the adapter design worth a look after the equations are corrected and code released. I would send this to peer review because a referee could confirm the typo and push a major revision, or catch a deeper mismatch. But I wouldn't cite it as is.","headline":"Strong benchmark numbers with a core fusion equation that cannot implement the claimed CNN-ViT complementation; possibly a typo, but no code and several internal inconsistencies make the central claim unverifiable.","tokens_in":13705,"tokens_out":2328,"would_cite":false,"duration_ms":27106,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A parallel CNN-ViT network with dynamic fusion sets new recall records on seven visual place recognition benchmarks.","keywords":["Visual place recognition","Local-global feature fusion","Dynamic feature fusion","Vision Transformer","CNN-ViT hybrid","Frequency-spatial adapter","Recall@1","Parameter-efficient fine-tuning"],"falsifier":"Re-run each baseline from Tables 2 and 3 under the exact GSV-Cities training split, query/database partition, and Recall@N metrics used here, and check whether the Recall@1 margins persist; if the margins shrink to near zero, the claim of significant outperformance fails. Alternatively, replace the DFM with fixed-ratio fusion of the same two streams; if Recall@1 does not drop materially on Pitts30k, SPED, and MSLS val, then the dynamic mechanism is not the cause of the gains.","tokens_in":12634,"feed_emoji":"🗺️","tokens_out":4349,"duration_ms":50817,"temperature":0.7,"pith_summary":"The paper proposes LGCN, a visual place recognition network that runs a CNN and a vision Transformer in parallel and fuses their features with a learned dynamic weighting instead of a static rule. It claims this local-global complementation lets a single descriptor capture both fine-grained texture and global scene context, overcoming a limitation of prior single-architecture methods. On benchmarks, the method reports state-of-the-art Recall@1 scores of 95.0% on Pitts30k, 95.5% on Pitts250k, 93.3% on SPED, and strong results on cross-season and cross-domain datasets. The practical significance, if true, is that frozen ViT backbones can be adapted with lightweight adapters and hybrid fusion to substantially improve robot localization and navigation accuracy without retraining a large model.","feed_headline":"CNN-ViT fusion hits 95% Recall@1 for place recognition","feed_subtitle":"Pairs local CNN details with global ViT context to beat prior state of the art on seven benchmarks.","key_machinery":"The two load-bearing mechanisms are the dynamic feature fusion module (DFM) and the frequency-spatial fusion adapter (FSA). DFM computes per-pixel, per-channel attention from the sum of the CNN and ViT feature maps using two 1x1 convolutions with ReLU and sigmoid, then recombines the ViT features with a learnable scaling factor so the network can emphasize different branches in different image regions. FSA is a lightweight adapter inserted into frozen ViT blocks: a frequency branch applies a 2D FFT, modulates the amplitude spectrum, and returns via inverse FFT, while a spatial branch uses depthwise convolutions; their outputs are added as a residual to the ViT output. The final descriptor is produced by feeding the fused features into a cross-image correlation-aware module, a component taken from the CricaVPR method.","core_discovery":"The paper claims that the best VPR descriptors come from jointly exploiting a CNN's local detail sensitivity and a ViT's global context, combined by a learned dynamic weighting rather than a fixed-ratio sum. Concretely, ResNet-50 features are upsampled to match a frozen DINOv2-pretrained ViT-B feature map, the two are added, and a bottleneck gating sub-network generates spatial-channel attention weights modulated by a learnable global scaling factor. In parallel, frequency-spatial adapters inserted after the ViT's attention layers modulate the amplitude spectrum in the frequency domain and apply depthwise convolutions in the spatial domain, adding local inductive bias without unfreezing the backbone. The method reports Recall@1 of 95.0% on Pitts30k, 95.5% on Pitts250k, 93.3% on SPED, 91.0% on MSLS val, and 89.8% on Nordland, with ablations attributing the gains to the adapter, the CNN stream, and the dynamic fusion module.","pith_inferences":["Editorial inference: Eq. (5) as written multiplies both recombined branches by the ViT feature map, so the CNN stream's contribution enters only through the summed feature F used to generate gating weights; the prose claims the fusion enhances CNN features in texture-rich regions, but the equations do not explicitly show a separate weighted CNN term. A reader should verify whether the intended fus","Editorial inference: The reported margins over prior work may be sensitive to evaluation protocol, since no baseline is re-run under the exact same training split and query/database partition; an independent re-implementation with identical protocol would test whether the gains hold.","Editorial inference: The frequency-spatial adapter idea could transfer to other retrieval tasks that need invariance to illumination and seasonal change, such as long-term visual SLAM or image-based re-localization under weather variation, which are not tested in this paper."],"forward_implications":["If LGCN is right, combining a frozen ViT with a CNN through learned dynamic fusion is a practical recipe for VPR accuracy, reaching 95.0% Recall@1 on Pitts30k and 95.5% on Pitts250k.","The ablation results imply each component is individually useful: adding FSA to the frozen ViT baseline improves Recall@1 by 6.2% on Pitts30k, and adding the CNN stream and DFM each give further gains.","Dynamic weighted fusion is claimed to outperform static concatenation and fixed-ratio summation, since the +DFM variant beats the +CNN Stream variant that uses simple concatenation.","The method's strong performance on SPED, Nordland, and St-Lucia suggests that frequency-domain adaptation and hybrid fusion generalize across seasonal, low-texture, and dynamic urban scenes.","Because the ViT backbone stays frozen and only lightweight adapters and fusion parameters are trained, the approach is parameter-efficient relative to full fine-tuning."],"supporting_citations":[{"why":"GSV-Cities is the training dataset used to fine-tune LGCN, so it anchors the training protocol for all comparisons.","marker":"[62]"},{"why":"Defines the Pittsburgh datasets (Pitts250k and Pitts30k) used for the main Recall@1 evaluation.","marker":"[56]"},{"why":"Provides the MSLS val dataset used to evaluate cross-city generalization and long-term appearance change.","marker":"[58]"},{"why":"Provides the SPED dataset with lighting and seasonal variation used to test robustness.","marker":"[57]"},{"why":"Provides the Nordland dataset with dramatic seasonal changes used to measure cross-season retrieval.","marker":"[59]"},{"why":"NetVLAD is a foundational CNN-based baseline whose aggregation approach informs many VPR descriptors and which LGCN must outperform.","marker":"[33]"},{"why":"MixVPR is a strong CNN-based baseline with a fully-connected feature-mixing architecture that LGCN compares against.","marker":"[37]"},{"why":"SelaVPR is a ViT-based baseline with adapter modules and re-ranking, representing the ViT-only approach LGCN extends.","marker":"[42]"},{"why":"CricaVPR is both a baseline and the source of the cross-image correlation-aware module used to generate LGCN's final descriptor.","marker":"[43]"},{"why":"The deep visual geo-localization benchmark supplies the evaluation protocol and standard dataset partitions used in the experiments.","marker":"[55]"}],"fun_headline_variants":["CNN-ViT fusion reaches 95% Recall@1 for place recognition","Hybrid CNN-ViT network attains 95% Recall@1 in VPR","Dynamic CNN-ViT fusion scores 95% Recall@1 on benchmarks","Local CNN plus global ViT yields 95% Recall@1 in VPR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains over prior work rest on the assumption that the baseline numbers in Tables 2 and 3 were produced under the same training data, evaluation splits, and re-ranking conditions as LGCN, since no baseline is re-run in this paper.","fun_headline_variants_meta":{"raw":{"variants":["CNN-ViT fusion reaches 95% Recall@1 for place recognition","Hybrid CNN-ViT network attains 95% Recall@1 in VPR","Dynamic CNN-ViT fusion scores 95% Recall@1 on benchmarks","Local CNN plus global ViT yields 95% Recall@1 in VPR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000657,"raw_usage":{"total_tokens":3019,"prompt_tokens":967,"completion_tokens":2052,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":1965}},"tokens_in":583,"tokens_out":2052,"duration_ms":19174,"temperature":1.0,"reasoning_tokens":1965,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:51:24.321582+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run each baseline from Tables 2 and 3 under the exact GSV-Cities training split, query/database partition, and Recall@N metrics used here, and check whether the Recall@1 margins persist; if the margins shrink to near zero, the claim of significant outperformance fails. Alternatively, replace the DFM with fixed-ratio fusion of the same two streams; if Recall@1 does not drop materially on Pitts30k, SPED, and MSLS val, then the dynamic mechanism is not the cause of the gains.","supporting_citations":[{"cited_title":"Neurocomputing513, 194–203 (2022)","cited_arxiv_id":null,"evidence_quote":"GSV-Cities is the training dataset used to fine-tune LGCN, so it anchors the training protocol for all comparisons."},{"cited_title":"In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"Defines the Pittsburgh datasets (Pitts250k and Pitts30k) used for the main Recall@1 evaluation."},{"cited_title":"In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"Provides the MSLS val dataset used to evaluate cross-city generalization and long-term appearance change."},{"cited_title":"In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"Provides the SPED dataset with lighting and seasonal variation used to test robustness."},{"cited_title":"In: Proceedings of Workshop on Long-term Autonomy, IEEE International Conference on Robotics and Automation, p","cited_arxiv_id":null,"evidence_quote":"Provides the Nordland dataset with dramatic seasonal changes used to measure cross-season retrieval."},{"cited_title":"In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"NetVLAD is a foundational CNN-based baseline whose aggregation approach informs many VPR descriptors and which LGCN must outperform."},{"cited_title":"In: Proceedings of the IEEE Winter Conference on Applications of Computer Vision, pp","cited_arxiv_id":null,"evidence_quote":"MixVPR is a strong CNN-based baseline with a fully-connected feature-mixing architecture that LGCN compares against."},{"cited_title":"In: Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"CricaVPR is both a baseline and the source of the cross-image correlation-aware module used to generate LGCN's final descriptor."},{"cited_title":"In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"The deep visual geo-localization benchmark supplies the evaluation protocol and standard dataset partitions used in the experiments."}],"review_version":1}