{"id":"8a09cd45-3030-4c1e-9028-a9b86f632434","arxiv_id":"2504.16433","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"FrogDogNet applies Fourier filtering and self-attention to CLIP visual features before prompt learning, reporting new state-of-the-art remote sensing domain generalization results on four benchmarks.","lead":"The paper proposes a new way to adapt CLIP, an AI model that connects images and text, to satellite and aerial imagery by keeping only the smooth, low-frequency parts of its internal image representation and discarding the rest. If the reported gains hold, it would give remote sensing teams a simple, low-cost recipe for making CLIP-based classifiers more robust across different satellites, seasons, and regions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Fourier Filter Block is never ablated, so the headline gains cannot be attributed to frequency filtering; a control with FFB removed is needed.","rationale":"The reader's weakest assumption was whether FFT truncation of the CLIP feature embedding corresponds to meaningful frequency content. My concern is closely related but more operational: even if the frequency premise were meaningful, the paper provides no experiment isolating the FFB, so the headline claim is not yet supported. The ablation studies in Section 4.2 test the RPA loss, context length, Lambda, backbone, and shot count, but not the presence or absence of the FFB. This is the single most load-bearing gap because the central claim is that Fourier frequency retention improves domain generalization. A control experiment with the FFB removed or replaced by a random mask would settle the attribution. The paper does report consistent gains across multiple tables and provides a code link, which are positive signs, but without the FFB ablation the gains could plausibly come from the projection, self-attention, or alignment loss components. I therefore do not change the reader's CONDITIONAL verdict, since the concern is addressable in revision and does not by itself demonstrate that the method fails.","tokens_in":20166,"tokens_out":3375,"duration_ms":36905,"concrete_test":"Run a single ablation set on the B2N 16-shot PatternNet configuration from Table 4 exactly as reported (lambda=0.3, Lambda=0.5, M=4, ViT-B/16, 50 epochs, three seeds), but replace the FFB with the identity map while keeping the projection, self-attention, Meta-Net, and RPA loss unchanged. If the harmonic mean stays within seed noise of 85.63, the FFB is not the driver of the improvement. As a second arm, replace the FFT low-pass with a fixed random 350-channel mask to test whether any hard threshold, rather than low-frequency structure, causes the gain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"FrogDogNet's core novelty is the Fourier Filter Block (FFB, Section 3.2.3, Figure 3), but the paper never ablates it. Section 4.2 ablates the RPA loss, context length, the balancing hyperparameter Lambda, the backbone, and the shot count, yet contains no condition with the FFB removed or replaced. The only evidence tying the FFB to the reported gains is Figure 1(b), a sensitivity sweep that selects k=350 using the same generalization tasks reported in Tables 1-3, with the supporting details deferred to a missing supplementary. Because the full pipeline also contains a projection network, self-attention, residual scaling by lambda, the Meta-Net, and the RPA loss, the 3.91-point average HM gain over StyLIP in Table 1 and the 5-7-point cross-dataset gains in Table 2 cannot be attributed to low-frequency retention without a direct control. Additionally, applying FFT along the 512-dimensional CLIP feature embedding has no obvious spatial-frequency interpretation; a fixed hard threshold of 350 coefficients may simply act as a regularizing channel mask. The non-monotonic shot curve in Table 4 (HM 85.63 at 16 shots dropping to 77.28 at 32 shots) further suggests that the reported operating point is not a stable optimum.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes FrogDogNet, a prompt-learning framework for CLIP-based remote sensing scene classification and domain generalization. The pipeline extracts CLIP visual features, refines them through a projection network and self-attention with residual connections, then passes them through a Fourier Filter Block (FFB) that retains the top k=350 low-frequency FFT coefficients of the 512-dimensional feature vector and applies an inverse FFT. A Meta-Net converts the filtered features into visual prompt tokens, and a Remote Sensing Prompt Alignment (RPA) loss aligns learned text prompts with RS-specific prompt initializations. Experiments on PatternNet, RSICD, RESISC45, and MLRSNet report state-of-the-art results on base-to-new class generalization, cross-dataset transfer, and single-source multi-target domain generalization, with an average base-to-new harmonic mean of 76.64 versus 72.73 for StyLIP.","tokens_in":20483,"tokens_out":5401,"duration_ms":49478,"significance":"If the reported gains are real and attributable to the FFB, the paper would make a useful contribution: it provides a comprehensive benchmark comparison across four RS datasets and three domain-generalization protocols, it includes a public code link, and its results are consistently above strong prompt-learning baselines such as StyLIP, APPLeNet, and MaPLe. The central hypothesis that low-frequency feature retention improves invariance to domain shift is interesting and testable. However, the current evidence does not yet isolate the FFB as the cause of the gains, the FFT-on-features mechanism lacks a clear semantic interpretation, and the reported results lack variance estimates and show non-monotonic behavior in the shot-response curve. These issues are fixable but need to be addressed before the central claim can be accepted.","major_comments":[{"comment":"The Fourier Filter Block is never ablated. Section 4.2 ablates the RPA loss, context length, Λ, backbone, and shot count, but no condition removes or replaces the FFB. Because the final model also contains the projection network, self-attention, residual scaling, Meta-Net, and RPA loss, the reported gains over StyLIP (3.91 average harmonic mean in Table 1, and 5.36–6.98 points in Table 2) cannot be attributed to frequency filtering. Please add an ablation with the FFB replaced by an identity mapping and by a fixed random mask of the same size, keeping everything else fixed.","section":"§4.2, Tables 1–3"},{"comment":"The conceptual basis of the FFB is underspecified. The DFT formulas in Eqs. (4)–(5) are given for image-like spatial inputs, but the block operates on the 512-dimensional CLIP visual embedding fv(x); it is not explained how low-frequency coefficients of this learned feature vector correspond to spatial or semantic frequency, why a hard threshold at k=350 is appropriate, or how the FFT is taken (real versus complex, and along which axis). The claim that low-frequency retention removes background artifacts and preserves class-relevant structure is an untested assumption; it is equally plausible that the operation acts as a fixed regularizing channel mask.","section":"§3.2.3, Figure 3"},{"comment":"The choice k=350 is made from a sensitivity analysis on the same generalization tasks and datasets used for the final comparison in Tables 1–3, with supporting details deferred to a missing supplementary. Selecting a hyperparameter on the evaluation tasks can inflate reported performance. Please provide a nested or held-out selection procedure, or report k sensitivity separately for each dataset and task so the reader can gauge the stability of the choice.","section":"Figure 1(b), §4"},{"comment":"The shot-response curve is non-monotonic: the harmonic mean rises to 85.63 at 16 shots and falls to 77.28 at 32 shots. This is surprising because more training data should not degrade performance, and it suggests that the reported operating point is not a stable optimum. Please investigate and report whether this drop persists with different seeds, schedulers, or the final hyperparameter settings used for Tables 1–3.","section":"Table 4"},{"comment":"All tables report averages over three random seeds but no standard deviations or confidence intervals. Given the small margins in several comparisons (for example, Table 1 RESISC45 base accuracy is 0.27 below StyLIP, and Table 3 source accuracy is 0.05 below StyLIP), the reader cannot assess whether the differences are significant. Please add variance estimates or per-seed results.","section":"§4.1, Tables 1–3"}],"minor_comments":[{"comment":"The argmin notation inside the definitions of Lce and Ltotal is not a loss definition; these equations should be written as minimization objectives over the parameters.","section":"§3.3, Eqs. (18) and (20)"},{"comment":"The expression Xfinal = Xout/X + X divides by a matrix without defining the operation; please specify elementwise division and check the dimensions.","section":"§3.2.2, Eq. (13)"},{"comment":"Several analyses are deferred to a supplementary that is not included in the arXiv v1 manuscript (Figure 1(b) details, dataset splits, and hyperparameter trends on the remaining datasets); please include the supplementary or move the essential details into the main text.","section":"§4.2"},{"comment":"The paper claims that full-image features introduce noise and background artifacts and that the FFB removes them, but no quantitative evidence directly supports this causal story beyond the t-SNE visualization; consider adding a direct comparison of filtered versus unfiltered features on a controlled noise or background perturbation.","section":"Abstract and §1"},{"comment":"Figure 1(a) illustrates frequency filtering on an image, whereas the FFB operates on the CLIP feature vector; the relationship between these two levels should be clarified to avoid confusing the reader about where the filtering actually happens.","section":"Figure 1(a) vs Figure 3"},{"comment":"The GFLOPS comparison is only against CoCoOp and APPLeNet, and reporting a percentage difference with two decimals for a 0.009 GFLOPS gap is misleading without noting that such differences are likely within measurement noise.","section":"Table 5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is not ready for acceptance in its current form. The missing FFB ablation is the central obstacle: without it, the paper's main novelty claim is unsupported. The selection of k=350 on the same evaluation tasks and the non-monotonic shot curve also raise concerns about the stability of the reported operating point. If the authors can supply the missing ablation, a random-mask control, variance estimates, and a clearer justification of the FFT-on-features mechanism, the paper may become publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my take: the paper deserves a serious referee, but the central claim needs a control.\n\nWhat's actually new: a prompt-learning pipeline for CLIP-based remote sensing DG that combines a Fourier Filter Block on the visual feature embedding, projection + self-attention refinement, and a remote-sensing prompt alignment loss. The building blocks are known, but the specific combination is not, and the reported numbers are consistently ahead of strong baselines: +3.91 harmonic mean over StyLIP in base-to-new, roughly +5 to +7 points in cross-dataset, and solid gains in multi-target DG. That is practically useful if it holds.\n\nWhat the paper does well: it evaluates on four standard RS datasets across three DG settings, ablates several design choices, reports compute, and the RPA loss clearly helps (Figure 6). The writing is direct and the related work is thorough.\n\nWhere it's soft: the Fourier Filter Block is never ablated. Section 4.2 removes RPA loss, changes context length, Lambda, backbone, and shot count, but no run drops or replaces the FFB. Without that control, the headline gains cannot be attributed to frequency filtering. The only evidence is Figure 1(b), a sensitivity sweep that picks k=350 using the same generalization tasks reported in the main tables, with details deferred to a missing supplementary. That is tuning on the test task, not evidence. Conceptually, applying an FFT to the 512-d CLIP feature vector has no clear spatial-frequency semantics; the hard threshold might be a regularizing channel mask. The shot curve is non-monotonic (HM 85.63 at 16 shots, 77.28 at 32), and there are no standard deviations despite three seeds. These are addressable in a revision.\n\nBottom line: a useful empirical recipe, but the mechanism is unproven. If the FFB ablation kills the gains, the paper is still a decent prompt-learning method; if the gains survive, it's a credible contribution. I would not cite it for the Fourier claim until that control exists.\n\nRecommendation: send it to peer review, with reviewers asked to require an FFB ablation and a held-out selection of k.","headline":"A useful prompt-learning recipe for remote sensing DG with consistently strong numbers, but the core Fourier-filtering claim is never directly tested.","tokens_in":21012,"tokens_out":3730,"would_cite":false,"duration_ms":34886,"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":"Filtering CLIP's image features to their strongest low-frequency Fourier components improves domain generalization in remote-sensing scene classification.","keywords":["remote sensing scene classification","domain generalization","prompt learning","vision-language models","CLIP","Fourier filtering","few-shot learning","frequency-domain features"],"falsifier":"Replace the Fourier Filter Block in the published pipeline with a random mask that keeps 350 of the 512 coordinates, with a PCA projection to 350 dimensions, and with a version where the FFT is applied after permuting the feature dimensions. If accuracy on the three domain-generalization tasks stays within noise across these variants, the specific frequency ordering is not doing the work.","tokens_in":19991,"feed_emoji":"🛰️","tokens_out":10541,"duration_ms":91719,"temperature":0.7,"pith_summary":"FrogDogNet claims that prompt learning for CLIP in remote sensing underperforms because it feeds the whole image embedding, including background clutter and sensor noise, into the prompt generator. The paper's remedy is to transform the refined visual embedding into the frequency domain, keep only the 350 largest low-frequency coefficients out of 512, and transform back before generating visual prompt tokens. Across four remote-sensing datasets and three domain-generalization tasks, the method reports consistent gains over prior prompt-learning baselines, including a base-to-new harmonic mean of 76.64 (versus 72.73 for the strongest prior method) and cross-dataset improvements of roughly five to seven accuracy points. The central claim is that low-frequency structure in CLIP's feature space carries class-relevant, domain-invariant information, and that prompt tuning should build on that structure rather than the full feature vector.","feed_headline":"Fourier-filtered CLIP features beat prior prompt learners","feed_subtitle":"Keeping 350 low-frequency image features lifts generalization across four remote-sensing datasets.","key_machinery":"The load-bearing object is the Fourier Filter Block (FFB), which applies a fast Fourier transform to the processed image embedding, keeps the top k=350 low-frequency coefficients out of 512, and applies an inverse transform to reconstruct a filtered embedding. A projection network and a scaled self-attention module with residual connections refine the raw CLIP feature before filtering, and a lightweight Meta-Net then converts the filtered embedding into M visual tokens that are added to learnable text tokens. A Remote Sensing Prompt Alignment loss, minimized jointly with cross-entropy, pulls the learned text prompts toward CLIP's own remote-sensing text embeddings. The FFB's job is to make the visual prompt tokens depend on class-redundant low-frequency structure, which the authors identify as the invariant content that survives domain shift.","core_discovery":"The paper's central discovery is that truncating the discrete Fourier transform of CLIP's processed image features to the top 350 of 512 low-frequency coefficients—instead of using the full feature vector—yields visual prompt tokens that generalize better across remote-sensing domains. The authors argue that full-image features carry within-class variation from backgrounds, sensors, and atmospheric conditions, and that these artifacts concentrate in the high-frequency components of the embedding. Removing them before a lightweight Meta-Net generates prompt tokens, together with a self-attention stage that preserves boundary-level local cues, produces a prompt-learning pipeline that outperforms prior methods on base-to-new class generalization, cross-dataset transfer, and single-source multi-target domain generalization. The reported cost of the added machinery is small: 192.361 GFLOPS, nearly identical to CoCoOp and below APPLeNet.","pith_inferences":["If the frequency hypothesis transfers, the same Fourier Filter Block recipe should improve other frozen vision-language encoders such as remote-sensing-specific CLIP variants; that is a direct test the paper does not run.","The cutoff of 350 out of 512 coefficients is tuned on these four datasets; a practical deployment would need to verify whether that single cutoff survives new sensors or whether per-domain tuning is required.","Because the prompt-alignment loss alone yields a 9.3% cross-dataset gain when prompts differ, an ablation that removes the FFB while keeping that loss would separate the contributions of frequency filtering from prompt alignment.","The t-SNE comparison suggests better class separation, but that visual claim could be quantified with linear-probe accuracy or nearest-centroid distances on the filtered embeddings."],"forward_implications":["If the reported numbers hold, the method reaches 76.64 average harmonic mean for base-to-new classes across the four datasets, 3.91 points above StyLIP and roughly 7.8 points above CoOp.","Cross-dataset transfer from PatternNet improves over StyLIP by 6.98, 6.41, and 5.36 points on RSICD, RESISC45, and MLRSNet, respectively.","Single-source multi-target generalization improves over StyLIP by 1.83, 2.34, and 5.04 points on the three target datasets while remaining within 0.05 points of StyLIP on the source.","Prompt alignment is particularly valuable when train and test prompts differ: adding the Remote Sensing Prompt Alignment loss gives a 9.3% average gain in the cross-dataset setup with mismatched prompts.","At 16 shots on PatternNet, the method reports an 85.63 harmonic mean, but its 32-shot result (77.28) is lower, so the benefit is not monotonic in the number of training examples."],"supporting_citations":[{"why":"CLIP is the frozen vision-language backbone whose encoders FrogDogNet adapts and the zero-shot baseline it must beat.","marker":"[81]"},{"why":"CoOp introduces learnable context prompts, the optimization route FrogDogNet extends and a primary comparison baseline.","marker":"[139]"},{"why":"CoCoOp supplies the Meta-Net prompt-generation design and the conditional-prompting baseline whose features FrogDogNet filters.","marker":"[138]"},{"why":"StyLIP is the strongest remote-sensing prompt-learning baseline; FrogDogNet's improvements are measured against it.","marker":"[7]"},{"why":"APPLeNet defines the remote-sensing few-shot evaluation protocol, the v2 dataset splits, and the multi-scale prompting contrast.","marker":"[95]"},{"why":"The fast Fourier transform provides the efficient frequency-domain operation that the Fourier Filter Block relies on.","marker":"[17]"},{"why":"Knowledge-guided context optimization motivates the Remote Sensing Prompt Alignment loss that pulls learned prompts toward RS text embeddings.","marker":"[33]"},{"why":"Self-attention is the mechanism that preserves local boundary cues before Fourier filtering.","marker":"[106]"}],"fun_headline_variants":["Keep 350 low-frequency CLIP features for better remote-sensing generalization","Fourier-trimmed CLIP features boost remote sensing domain shift","Low-freq CLIP filter sharpens prompt learning in RS","Dropping high-frequency CLIP noise improves domain generalization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that cutting a CLIP feature vector's Fourier transform down to its 350 largest low-frequency coefficients keeps class-relevant structure and discards noise and background artifacts; if an embedding's FFT has no meaningful frequency ordering, the filter is not the source of the gains.","fun_headline_variants_meta":{"raw":{"variants":["Keep 350 low-frequency CLIP features for better remote-sensing generalization","Fourier-trimmed CLIP features boost remote sensing domain shift","Low-freq CLIP filter sharpens prompt learning in RS","Dropping high-frequency CLIP noise improves domain generalization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000291,"raw_usage":{"total_tokens":1694,"prompt_tokens":934,"completion_tokens":760,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":686}},"tokens_in":550,"tokens_out":760,"duration_ms":7402,"temperature":1.0,"reasoning_tokens":686,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:03:46.890858+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the Fourier Filter Block in the published pipeline with a random mask that keeps 350 of the 512 coordinates, with a PCA projection to 350 dimensions, and with a version where the FFT is applied after permuting the feature dimensions. If accuracy on the three domain-generalization tasks stays within noise across these variants, the specific frequency ordering is not doing the work.","supporting_citations":[{"cited_title":"Learn- ing transferable visual models from natural language super- vision","cited_arxiv_id":null,"evidence_quote":"CLIP is the frozen vision-language backbone whose encoders FrogDogNet adapts and the zero-shot baseline it must beat."},{"cited_title":"Learning to prompt for vision-language models","cited_arxiv_id":null,"evidence_quote":"CoOp introduces learnable context prompts, the optimization route FrogDogNet extends and a primary comparison baseline."},{"cited_title":"Conditional prompt learning for vision-language models","cited_arxiv_id":null,"evidence_quote":"CoCoOp supplies the Meta-Net prompt-generation design and the conditional-prompting baseline whose features FrogDogNet filters."},{"cited_title":"StyLIP: Multi-Scale Style-Conditioned Prompt Learning for CLIP-based Domain Generalization","cited_arxiv_id":"2302.09251","evidence_quote":"StyLIP is the strongest remote-sensing prompt-learning baseline; FrogDogNet's improvements are measured against it."},{"cited_title":"Applenet: Visual attention parameterized prompt learning for few-shot remote sens- ing image generalization using clip","cited_arxiv_id":null,"evidence_quote":"APPLeNet defines the remote-sensing few-shot evaluation protocol, the v2 dataset splits, and the multi-scale prompting contrast."},{"cited_title":"Attention is all you need","cited_arxiv_id":null,"evidence_quote":"Self-attention is the mechanism that preserves local boundary cues before Fourier filtering."}],"review_version":1}