{"id":"decec3d2-cbb4-43fa-870e-7a7bbb34096b","arxiv_id":"2412.00784","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A decoder transformer with learnable queries plus a low-rank parallel adapter for frozen DINOv2 achieves state-of-the-art visual place recognition on multiple benchmarks with reduced training memory.","lead":"This paper proposes a new feature-aggregation model, EDTformer, that turns image features into a global descriptor using learnable queries and a lightweight parallel adaptation of a frozen DINOv2 backbone. The authors report state-of-the-art retrieval accuracy on multiple visual place recognition benchmarks with lower training memory use than previous methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline comparability is the load-bearing risk: BoQ is re-run with a modified image size and no code is released, so the claimed SOTA margins may not reflect a controlled comparison.","rationale":"I focused on baseline comparability rather than memory efficiency because the memory claim is supported by a direct mechanism (no backpropagation through the backbone, Eq. 10) and by the ablation in Table VI. The decoder architecture is clearly described and the LoPA memory advantage is plausible. The weakness is that the empirical SOTA claim, which is the headline, depends on numbers from heterogeneous sources and a modified BoQ baseline. The footnote is an explicit self-acknowledged protocol change, and the lack of code makes it untestable. This does not mean the result is wrong; it means the paper is conditional on releasing artifacts and running a unified evaluation. The reader's weakest_assumption identified the same comparability issue, so I agree. The selection-on-test-set issue compounds it, but the controlled comparison is the primary blocker. The recommended verdict remains CONDITIONAL, with no change from the reader's verdict.","tokens_in":24165,"tokens_out":11044,"duration_ms":110666,"concrete_test":"Run BoQ with the official code and its original image-size setting (both train and inference as in the BoQ paper), and run EDTformer with the same training set, image sizes, and evaluation code, on at least MSLS-challenge, Tokyo24/7, Pitts30k, and AmsterTime. Report R@1 over 5 training seeds. If official BoQ numbers exceed the paper's BoQ‡ numbers, or if the EDTformer margins over BoQ/SelaVPR fall within seed noise on a majority of datasets, the SOTA claim is not established. As a secondary check, select L, M, and rank on a held-out validation split (e.g., a random subset of GSV-Cities) and confirm the margins survive.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that EDTformer outperforms existing single-stage and two-stage VPR methods. The strongest evidence for this is the margin over BoQ on MSLS-challenge (78.4 vs 75.9 R@1), Tokyo24/7 (97.1 vs 95.2), and over SelaVPR on AmsterTime (65.2 vs 53.6). These comparisons are only valid if the baseline numbers are measured under equivalent protocols. The paper itself flags the BoQ row as a re-run “strictly following its training pipeline, except for keeping the same image size for both training (224×224) and inference (322×322) as our method” (footnote, Table II). Because neither code nor checkpoints are released, this modified BoQ cannot be checked; if the official BoQ uses a different inference size, the reproduced number may not be the BoQ that the community would report. All other baselines are taken from the original papers, which use different training pipelines (e.g., SALAD and SelaVPR may use different backbones or fine-tuning strategies), so the table mixes protocols. On top of this, the L=2, M=64, rank=4 hyperparameters are selected using ablations on MSLS-val, SPED, Pitts30k, and Nordland, the same datasets later used for headline results; this selection-on-test-data bias further inflates the reported margins. The claim that EDTformer beats two-stage re-ranking methods rests on a single unverified comparison rather than a controlled head-to-head.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EDTformer, a transformer-decoder-based feature aggregator for single-stage visual place recognition (VPR), combined with LoPA, a low-rank parallel adaptation module that keeps a DINOv2 backbone frozen while refining its intermediate features. The global descriptor is formed by a set of learnable queries passed through stacked simplified decoder blocks (self-attention and cross-attention only) followed by two linear layers. The authors report state-of-the-art Recall@N on eight benchmarks, including MSLS, Tokyo24/7, SPED, Nordland, AmsterTime, and SVOX, and claim lower training memory and inference time than comparable DINOv2-based methods. Internal ablations cover the aggregator choice, the adaptation method, the number of decoder blocks and queries, the removal of the FFN, the LoPA rank, and the transferability of LoPA to other aggregators.","tokens_in":24455,"tokens_out":3094,"duration_ms":31113,"significance":"If the empirical claims hold, the paper makes two useful contributions: a decoder-style query-based aggregation that is simpler and more parameter-efficient than BoQ, and a parallel adaptation approach that avoids backpropagation through a frozen foundation model, with a plausible memory advantage over adapter-based PETL. The internal ablations in Tables V-X are broadly consistent and support the main design decisions, and the demonstration that LoPA improves GeM, NetVLAD, and SALAD as well as EDTformer is a genuinely useful result. The central limitation is that the headline SOTA comparison is not a controlled experiment: the only directly re-run baseline (BoQ) is re-run under a modified image-size protocol, several other baselines are quoted from papers with different backbones and training pipelines, and structural hyperparameters are selected on the same evaluation sets later used for the headline tables. These issues affect the claimed margin over prior methods but not the internal consistency of the proposed architecture.","major_comments":[{"comment":"The headline claim that EDTformer outperforms BoQ and other SOTA methods is not supported by a controlled comparison. The footnote to Table II states that BoQ is re-run 'strictly following its training pipeline, except for keeping the same image size for both training (224x224) and inference (322x322) as our method.' Since the official BoQ release is not used and no code or checkpoints are provided, the reader cannot verify that this reproduced BoQ is the configuration the community would report; changes in inference resolution can meaningfully alter retrieval accuracy. Other baselines (e.g., SelaVPR with DINOv2-large, SALAD, Patch-NetVLAD, R2Former) are taken from their original papers with different training protocols. The margins cited as evidence (MSLS-challenge 78.4 vs 75.9, Tokyo24/7 97.1 vs 95.2, AmsterTime 65.2 vs 53.6) are therefore not attributable to the proposed method until all methods are evaluated under a common protocol, ideally with released code and checkpoints.","section":"Section IV-C, Table II and Table III"},{"comment":"The structural hyperparameters L=2, M=64, and rank=4 are selected by ablation studies on MSLS-val, SPED, Pitts30k, and Nordland (Tables VII, VIII, X), and the same datasets are then used for the headline results in Tables II and III. This selection-on-test-data procedure can inflate the reported margins and makes the comparison to methods whose hyperparameters were not tuned on these benchmark sets unfair. The authors should either use a separated validation set for model selection or explicitly quantify the sensitivity of the final results to the chosen hyperparameters.","section":"Section IV-D and Section IV-B"},{"comment":"No error bars, multiple seeds, or significance tests are reported for any of the benchmark numbers. Several of the claimed improvements are small in percentage terms (e.g., 1.5% R@1 over BoQ on Pitts30k, 0.2% over SALAD on MSLS-val), and without repeated runs or a statistical test it is unclear whether these differences are meaningful rather than run-to-run noise. At minimum, the authors should report the mean and standard deviation over at least three training runs for the main comparing methods and for their own method.","section":"Tables II and III"},{"comment":"The claim that EDTformer 'outperforms two-stage VPR methods which add a re-ranking with considerable cost' rests entirely on quoted numbers from Patch-NetVLAD, R2Former, and SelaVPR in Tables II and III. These methods are not re-run under the same training data, backbone, or evaluation pipeline, and the comparison does not report the candidate-set size or re-ranking cost used for the quoted two-stage numbers. Since the stated advantage over two-stage methods is a central selling point of the paper, the authors should provide a controlled head-to-head evaluation (same training data, same query/database splits, same evaluation script) or substantially soften the claim.","section":"Section I and Section IV-C, two-stage comparison"},{"comment":"The training-memory comparison in Table IV is measured with batch size 36, while the adaptation-method comparison in Table VI is measured with batch size 72. The paper should state why different batch sizes are used and confirm that the relative memory savings are not an artifact of the batch-size choice; ideally, all memory comparisons should be reported at the same batch size or with a scaling analysis.","section":"Section IV-B and Section IV-C"}],"minor_comments":[{"comment":"The text says 'PartialTuning used in BoQ [24] and SALAD [24]' but both citations point to SALAD; the BoQ reference should be [5] and the SALAD reference should be [24].","section":"Section IV-D, Table VI"},{"comment":"The row labeled 'SelaVPR [24]' should cite [4], not [24]; the same typo appears in the caption context of Table III.","section":"Table III"},{"comment":"The legend contains the typo 'EngenPlaces'; it should read 'EigenPlaces'.","section":"Fig. 1"},{"comment":"The method 'SFRS' is misspelled as 'SRFS' in Tables II and III; the reference list and Section IV use SFRS.","section":"Tables II and III"},{"comment":"The dimensions in Eq. (9) are not fully explained; in particular, how W3 acts on the transposed matrix to produce the final flattened descriptor should be stated explicitly in the text.","section":"Eq. (9)"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid internal story: the architecture is simple, the LoPA idea is plausible and its independence from the aggregator is a nice result, and the ablation tables are internally consistent. The main risk is that the SOTA comparison is not reproducible as described, especially because the only re-run baseline uses a modified image size and no code is released. I would recommend that the editor request the authors to either release code and checkpoints, re-run all baselines under a common protocol, or clearly restrict the claims to a controlled subset of the comparison. The current version is too dependent on unverifiable numbers for a definitive acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The paper is a solid empirical VPR contribution with a genuinely useful memory-saving idea (LoPA) and a clean decoder-based aggregation story. The headline SOTA margins, though, rest on a baseline comparison that isn't fully controlled, so I'd treat the exact numbers as provisional until someone checks the BoQ re-run.\n\nWhat's new: using a pure decoder with a single set of learnable queries for VPR aggregation isn't in BoQ, which uses encoder-style aggregation with multiple query sets. LoPA is a close relative of side-tuning/parallel PETL methods (LST, UniPT, etc.; the paper cites them), but applying it to DINOv2 for VPR and showing it works with other aggregators is a nice practical contribution. The memory and parameter numbers are the most persuasive part: 5.72 GB training memory vs 15.21 for BoQ, 0.08M trainable parameters, and Table XI shows LoPA lifts GeM, NetVLAD, and SALAD on multiple datasets. The internal ablations are consistent and support the design choices. The paper also cites the prior parallel-tuning work honestly, so no circularity concern.\n\nThe soft spot is the benchmark comparison. The BoQ row in Table II is a re-run with a modified inference image size (322x322 instead of whatever official BoQ uses), and since no code or checkpoints are released, that number can't be verified. Other baselines come from their original papers with different training pipelines, so the table mixes protocols. On top of that, the architecture hyperparameters (L=2, M=64, rank=4) are selected by ablations on the same test sets used for the headline results, which can inflate the margins. There are also no error bars or repeated runs. These are real methodological concerns, but they don't kill the core contribution: the memory savings and the decoder aggregation are structural, and the ablations are internally coherent.\n\nWho should read it: VPR people and anyone interested in parameter/memory-efficient adaptation of foundation models. It's worth a serious referee, but the referee should ask for code/checkpoints and a cleaner protocol—either use the official BoQ numbers or run all methods under one pipeline. I'd accept it for review with that request; if the authors can't provide the artifacts, the SOTA claim stays weak but the method itself is still interesting.","headline":"Solid empirical VPR paper with a genuinely useful memory-saving adaptation scheme (LoPA); the headline SOTA margins rest on a baseline comparison that needs verification before trusting the exact numbers.","tokens_in":25023,"tokens_out":3727,"would_cite":true,"duration_ms":33160,"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":"A decoder-only feature aggregator over a frozen DINOv2 backbone produces global place descriptors that outperform both single-stage and two-stage visual place recognition methods on multiple benchmarks, with roughly one-third the training…","keywords":["visual place recognition","feature aggregation","transformer decoder","learnable queries","low-rank parallel adaptation","parameter-efficient transfer learning","DINOv2","global descriptor"],"falsifier":"Retrain EDTformer and the leading baselines, especially BoQ and SALAD, under one identical protocol, same image size, crop, batch, epochs, and a validation split fixed before any hyperparameter choice, and evaluate on a held-out VPR benchmark not used in ablations; if the reported margins vanish or reverse, the central claim fails. A cheaper falsifier is to measure peak training GPU memory for all four methods at the same batch size, since the paper reports memory at batch 36 while ablations use batch 72.","tokens_in":23931,"feed_emoji":"📍","tokens_out":8099,"duration_ms":69397,"temperature":0.7,"pith_summary":"This paper tries to establish that a transformer decoder, used purely as a feature aggregator over a frozen foundation-model backbone, can produce place-recognition descriptors that beat both single-stage global-retrieval methods and two-stage methods that add expensive re-ranking. The authors build EDTformer from stacked simplified decoder blocks, self-attention and cross-attention without the feedforward network, where deep features act as keys and values and a set of 64 learnable queries is progressively refined, then compressed through two linear layers into a 4096-dimensional L2-normalized descriptor. A parallel low-rank adapter, LoPA, refines the intermediate features of a frozen DINOv2 backbone so that backpropagation never passes through the backbone, cutting training memory to roughly 5.72 GB. If correct, the practical upshot is that one-stage global retrieval can be both more accurate and considerably cheaper than the current two-stage state of the art.","feed_headline":"Decoder-only descriptor beats reranking VPR systems","feed_subtitle":"Global-only retrieval with 64 learnable queries beats re-ranking methods and trains in 5.72 GB.","key_machinery":"The load-bearing mechanism is the simplified decoder block: a self-attention layer followed by a cross-attention layer, with no feedforward network. A fixed set of 64 learnable queries enters the first self-attention layer; in cross-attention the queries attend to the deep features $F$ supplied as keys and values, and after $L=2$ stacked blocks the refined queries are compressed by two linear layers into a 4096-dimensional global descriptor. The supporting mechanism is LoPA (Low-rank Parallel Adaptation): a lightweight parallel network whose adaptation function $h(x)=s\\,W_u\\,\\sigma(W_d x)+x$ takes as input the sum of the previous LoPA output and the current intermediate feature of the frozen DINOv2 block, progressively refining features so that training gradients never backpropagate through the backbone. LoPA contributes only 0.08 million trainable parameters and about 1.74 GB of extra training memory.","core_discovery":"The paper's central claim is that decoder-style aggregation is what has been missing from single-stage VPR. In EDTformer, the image features extracted by the frozen (LoPA-adapted) DINOv2 backbone are linearly projected and then used as keys and values in cross-attention; the queries are learnable parameters, not image-dependent tokens. After two simplified decoder blocks, the learned queries hold the aggregated contextual information and are passed through two fully connected layers, one for dimensionality reduction and one to adjust the number of queries, before flattening and L2 normalization yields the global representation. The authors report that this 4096-dim descriptor achieves state-of-the-art Recall@1 on MSLS-challenge (78.4%), Tokyo24/7 (97.1%), SPED (92.4%), Nordland (88.3%), AmsterTime (65.2%), and the SVOX subsets, exceeding single-stage methods and also the two-stage SelaVPR on Nordland and MSLS, while training in about one-third to half the memory of comparable DINOv2-based aggregators.","pith_inferences":["We infer that the same decoder aggregator should transfer to other image-retrieval tasks where a compact global descriptor is desired, since nothing in EDTformer is place-specific beyond the training data.","The attention visualizations suggest each query functions as a soft landmark detector; a testable extension is to use the learned query-attention maps as weak labels for unsupervised landmark discovery.","If the memory savings really come from avoiding backpropagation through the backbone, LoPA-style parallel adapters could reduce training cost for other foundation-model transfer tasks; this is our inference, since the paper only demonstrates VPR.","A cleaner stress test of the reported margins would train BoQ and EDTformer under one identical protocol and would reserve the ablation datasets for a separate validation stage; we infer the margins could shift under such a protocol, though the paper's BoQ rerun already addresses image size."],"forward_implications":["A purely global descriptor can outperform two-stage re-ranking pipelines on Nordland and MSLS, so the extra cost of geometric verification or local-feature re-ranking is not required for state-of-the-art accuracy on those benchmarks.","Training memory of about 5.72 GB means DINOv2-based VPR can be fine-tuned on a single consumer GPU, widening access to foundation-model-based place recognition.","The 4096-dim descriptor can be shrunk to 512 dims by changing only the last linear layer while still beating several published global descriptors, which matters for large-database storage and search speed.","Because LoPA improves GeM, NetVLAD, and SALAD as well as EDTformer, the adaptation method is a drop-in efficiency upgrade for existing aggregators.","Consistent gains on day/night, seasonal, and domain-shift datasets indicate the decoder queries learn to attend to condition-invariant landmark regions."],"supporting_citations":[{"why":"BoQ is the main single-stage competitor that also uses learnable queries; the paper reproduces BoQ under matched image size and compares accuracy, descriptor dimension, and speed.","marker":"[5]"},{"why":"SALAD is a DINOv2-based single-stage baseline that EDTformer must beat, and it supplies the optimal-transport aggregation contrast and the partial-tuning baseline.","marker":"[24]"},{"why":"CricaVPR provides a DINOv2 adapter baseline (MultiConvAdapter) and cross-image encoder results compared in Tables II and VI.","marker":"[3]"},{"why":"SelaVPR is the two-stage baseline with local-feature re-ranking; its adapter and partial-tuning variants are compared for accuracy and memory.","marker":"[4]"},{"why":"DINOv2 is the frozen backbone whose intermediate features LoPA refines, and the whole pipeline builds on its pretrained representations.","marker":"[30]"},{"why":"Supplies the central memory argument that parameter-efficient adapters still require backpropagation through the frozen backbone, motivating the parallel LoPA design.","marker":"[37]–[39]"},{"why":"GSV-Cities is the training dataset whose place labels are used with the multi-similarity loss to train EDTformer.","marker":"[66]"}],"fun_headline_variants":["64 learnable queries beat two-stage VPR reranking","Decoder-only transformer beats reranking VPR methods","Single-stage VPR with learnable queries beats reranking","Learnable query aggregation beats two-stage VPR","Efficient decoder transformer beats reranking in VPR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline comparisons assume that recall numbers reported by different methods are directly comparable, since most baselines are taken from their original papers with different training pipelines, BoQ is re-run with a modified image size, and the architecture choices (two blocks, 64 queries, rank 4) are selected using ablations on the same benchmark sets that later produce the headline results.","fun_headline_variants_meta":{"raw":{"variants":["64 learnable queries beat two-stage VPR reranking","Decoder-only transformer beats reranking VPR methods","Single-stage VPR with learnable queries beats reranking","Learnable query aggregation beats two-stage VPR","Efficient decoder transformer beats reranking in VPR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001447,"raw_usage":{"total_tokens":5892,"prompt_tokens":1069,"completion_tokens":4823,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":685,"completion_tokens_details":{"reasoning_tokens":4747}},"tokens_in":685,"tokens_out":4823,"duration_ms":32463,"temperature":1.0,"reasoning_tokens":4747,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:00:34.967488+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain EDTformer and the leading baselines, especially BoQ and SALAD, under one identical protocol, same image size, crop, batch, epochs, and a validation split fixed before any hyperparameter choice, and evaluate on a held-out VPR benchmark not used in ablations; if the reported margins vanish or reverse, the central claim fails. A cheaper falsifier is to measure peak training GPU memory for all four methods at the same batch size, since the paper reports memory at batch 36 while ablations use batch 72.","supporting_citations":[{"cited_title":"Cricavpr: Cross-image correlation-aware representation learning for visual place recognition,","cited_arxiv_id":null,"evidence_quote":"CricaVPR provides a DINOv2 adapter baseline (MultiConvAdapter) and cross-image encoder results compared in Tables II and VI."},{"cited_title":"Towards seamless adaptation of pre-trained models for visual place recognition,","cited_arxiv_id":null,"evidence_quote":"SelaVPR is the two-stage baseline with local-feature re-ranking; its adapter and partial-tuning variants are compared for accuracy and memory."}],"review_version":1}