{"id":"2f064ae3-e200-4698-b556-76f750ae30ec","arxiv_id":"1908.01962","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"REAPS combines class-activation-map region cropping with a bidirectional LSTM over serialized spatial features to learn implicit part representations for fine-grained image classification.","lead":"This paper proposes REAPS, a fine-grained image recognition method that crops the object region using class activation maps, then feeds a serialized feature sequence through a bidirectional LSTM to learn implicit part representations. The authors report state-of-the-art accuracy on three benchmarks, though an omitted recent baseline weakens that claim.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is undermined by the omitted NTS-Net baseline: NTS-Net (ECCV 2018) reports 93.9% on Stanford Cars and 87.5% on CUB-200-2011 with image-level labels, both above REAPS+'s 93.5% and 86.8%.","rationale":"The decisive issue is not internal consistency of the equations but the external completeness of the comparison. The paper's own Table 3 sets the state-of-the-art bar at MA-CNN, yet NTS-Net (ECCV 2018) is a prior weakly supervised method with published numbers above REAPS+ on two of the three benchmarks. This directly undermines the abstract claim. I considered the reader's CAM-threshold assumption as an alternative; it is a real risk, but the reported accuracies show the pipeline extracts useful signal, and without an ablation no concrete failure is established. The omitted baseline is a concrete, checkable correctness issue, so it is more load-bearing for the stated central claim. The reader's explicit weakest_assumption is about the crop, so agreement is only partial, though the reader's rationale already correctly flags NTS-Net. Since the appropriate remedy is exactly the reader's CONDITIONAL revision (drop the unqualified SOTA claim, add the missing baseline, release code), the verdict need not change.","tokens_in":8648,"tokens_out":6852,"duration_ms":69238,"concrete_test":"Add a row for NTS-Net to Table 3 using its published numbers from the ECCV 2018 paper, or run the released NTS-Net code with the same VGG-19 backbone and 448×448 inputs used for REAPS+ and report the three accuracies. If NTS-Net's Stanford Cars accuracy is ≥93.5 or its CUB-200-2011 accuracy is ≥86.8, then REAPS+ is not state of the art on those benchmarks and the abstract's claim must be revised to \"competitive\" or restricted to methods it actually outperforms.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in the abstract and §4.2 is that REAPS/REAPS+ \"achieves the state of the art\" using only image-level labels. Table 3 omits NTS-Net (ECCV 2018), a directly relevant weakly supervised fine-grained recognition method published before this work. NTS-Net reports 93.9% on Stanford Cars, 91.4% on FGVC-Aircraft, and 87.5% on CUB-200-2011 with the same supervision. Against REAPS+ (93.5, 92.6, 86.8), NTS-Net is higher on Cars by 0.4 and on CUB by 0.7, and lower on Aircraft by 1.2. The text designates MA-CNN as the state of the art, but NTS-Net had already surpassed MA-CNN on all three benchmarks. Therefore the SOTA conclusion is not merely under-hedged; it is false on two of three benchmarks under the published comparison. This concern is independent of the CAM-threshold risk identified by the reader, which is real but unquantified; the omitted baseline is a direct factual gap that can be checked without re-running the experiments.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes REAPS, a weakly supervised fine-grained recognition framework that combines a CAM-based region attending network (RAN) with a part sequence-learning network (PSN). RAN computes a class activation map (Eq. 1), thresholds it to crop and zoom in on an attended region (Eq. 2), and PSN serializes the backbone features into N vectors, maps them to M implicit part representations with a bi-directional LSTM, and classifies with three combined softmax losses (Eq. 5). Experiments on Stanford Cars, FGVC-Aircraft, and CUB-200-2011 report REAPS+ accuracies of 93.5%, 92.6%, and 86.8%, respectively, and the paper claims state-of-the-art performance with only image-level labels.","tokens_in":8957,"tokens_out":5109,"duration_ms":52156,"significance":"The conceptual contribution is attractive: replacing fixed hard part detectors with an implicit soft-part sequence model on a CAM-cropped object region is simple, end-to-end trainable, and requires no part annotations. The paper evaluates on three standard benchmarks with a consistent backbone, and the qualitative visualizations support the claim that the part branch produces more compact feature maps. However, the central state-of-the-art claim is not supported by the published comparison: the directly comparable weakly supervised method NTS-Net (ECCV 2018) reports 93.9% on Stanford Cars, 91.4% on FGVC-Aircraft, and 87.5% on CUB-200-2011 with image-level labels, so REAPS+ is higher only on Aircraft. In addition, key hyperparameters needed for reproduction are not specified. The paper contains no machine-checked proofs or released code, so its value rests on the completeness and correctness of the empirical comparison, and that comparison is currently incomplete.","major_comments":[{"comment":"Table 3 and the discussion in Section 4.2 omit NTS-Net (Yang et al., ECCV 2018), a directly comparable weakly supervised fine-grained recognition method that reports 93.9% on Stanford Cars, 91.4% on FGVC-Aircraft, and 87.5% on CUB-200-2011 using only image-level labels. Under this published comparison, REAPS+ (93.5%, 92.6%, 86.8%) is not state of the art on two of the three benchmarks, and the sentence in Section 4.2 that REAPS+ \"obtains the best performance on three datasets\" is not supportable. The authors must add NTS-Net and any other comparable published weakly supervised baselines, and revise the state-of-the-art claims accordingly.","section":"Table 3 / Sec. 4.2"},{"comment":"Key hyperparameters of the proposed architecture are never specified: the sequence length N in Eq. (3), the implicit part count M in Eq. (4), the LSTM hidden size U, the number of LSTM layers, and the scaling factor used for the second PSN in REAPS+. Section 4.1 states only the loss weights lambda_1=lambda_2=lambda_3=1, the threshold tau=0.1, and the learning-rate schedule. Without these values the experiments cannot be reproduced, and the sensitivity of the reported margins to these choices cannot be assessed.","section":"Sec. 3.3 / Sec. 4.1"},{"comment":"The central design assumption is that the thresholded CAM crop in Eq. (2) preserves all diagnostic fine-grained parts before the part sequence model sees them, but the paper does not ablate tau or quantify the information loss caused by cropping. If the class activation map highlights only the most salient patch and drops other diagnostic regions, the crop removes information before PSN can model it; an ablation over tau values and a comparison of REAPS with and without cropping would be needed to substantiate the claim in Section 4.2 that RAN \"effectively depress[es] background noise\" without sacrificing fine-grained details.","section":"Eq. (2) / Sec. 4.2"},{"comment":"The reported differences are small on two of the three benchmarks (93.5 vs. 92.8 for MA-CNN and 93.9 for NTS-Net on Cars; 86.8 vs. 86.5 for MA-CNN and 87.5 for NTS-Net on CUB), yet no standard deviations, confidence intervals, or significance tests are reported. Given the missing baseline and unspecified hyperparameters, the conclusion that REAPS \"outperforms the existing methods\" is not robust; at minimum the authors should report the variance over multiple runs or justify single-run reporting.","section":"Table 3 / Sec. 4.2"}],"minor_comments":[{"comment":"The statement that REAPS+ surpasses MA-CNN \"by a relative 2.7% gain\" on FGVC-Aircraft is arithmetically inaccurate: 92.6% vs. 89.9% is a 2.7 percentage-point improvement and a 3.0% relative improvement.","section":"Sec. 4.2"},{"comment":"The text says the part branch gives a \"relative performance gain of 1.0%\" for PSN on Stanford Cars, but the numbers 91.3% and 92.3% differ by 1.0 percentage point; the wording should be corrected.","section":"Sec. 4.2 / Table 2"},{"comment":"There is an apparent inconsistency between the notation in Eq. (4), where the mapped part representation is [P_1,...,P_M] with M <= N, and the later sentence stating that the concatenated LSTM hidden states give P_P in R^{N x U}; the authors should clarify whether M=N in their implementation or describe the additional reduction step.","section":"Sec. 3.3 / Eq. (4)"},{"comment":"The baseline named \"REAPS wo PSN\" is not clearly defined in the text; it should be stated explicitly which branches are removed and how this differs from the \"PSN wo part\" model in Table 2.","section":"Table 3"},{"comment":"The abstract and conclusion claim \"state of the art performance\" without qualification; given the omitted NTS-Net comparison, this wording should be revised to a more defensible claim such as \"competitive with current methods\" unless the comparison is updated.","section":"Abstract / Sec. 5"},{"comment":"The notation U and M appears in Fig. 3 without definition in the caption or main text; please define all symbols at first use.","section":"Sec. 3.3 / Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The omission of NTS-Net is the most consequential issue: it is a well-known ECCV 2018 weakly supervised method with higher reported accuracy than REAPS+ on two of the three datasets, so the central SOTA claim is factually incorrect as written. Before any decision, I recommend that the editor verify the NTS-Net numbers against the original publication and require the authors to include the baseline, specify all hidden hyperparameters, and either add variance estimates or soften the comparative claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a competently written, plausible piece of incremental work. The new bit is the 'soft-part' idea: instead of detecting a fixed number of parts, serialize the spatial feature vectors and run a bidirectional LSTM to learn implicit part representations. That is a neat framing and the end-to-end training with three losses is sensible. The experiments cover the three standard fine-grained datasets and the reported numbers are competitive with several baselines. The paper clearly explains the motivation and the method, and the ablation showing the part branch helps is useful.\n\nThe problems are in the comparison and the claims. Table 3 omits NTS-Net (ECCV 2018), which is directly relevant and was published before this work. NTS-Net reports 93.9% on Stanford Cars and 87.5% on CUB, both above REAPS+ (93.5 and 86.8). The stress-test note is right: the abstract and Section 4.2 say 'state of the art', and on two of three benchmarks that is simply false under the published comparison. The text also says REAPS+ surpasses MA-CNN by a relative 2.7% on Aircraft, but on CUB REAPS (86.0) is below MA-CNN (86.5), so the overall superiority claim is undercut. This is not a matter of heavy hedging; it is a factual gap that any careful reader can check.\n\nSecond, the paper never specifies the sequence length N, the implicit part count M, or the LSTM hidden size U. Those are core hyperparameters for reproducing the method. The CAM threshold tau is set to 0.1 but there is no ablation on it, so the concern that the attended region may drop discriminative details is real but unquantified. No error bars are given, which is common but still a limitation when the margins are under one percent. The claim that REAPS is simple and robust is therefore not fully substantiated.\n\nThe idea itself is legitimately a combination of known ingredients, and the paper does not hide that; it cites CAM and RA-CNN. The novelty is in the combination, not in the components. That is enough for a workshop or a solid conference paper with careful revision.\n\nWho should read this? People working on weakly supervised fine-grained recognition will find the sequence-modeling approach worth knowing, especially the ablations. It should go to peer review, not be desk-rejected, but it needs a major revision: add NTS-Net, specify all hyperparameters, report error bars if possible, and tone down the SOTA claim to 'competitive'. As it stands, the evidence supports the method working reasonably well, not setting a new state of the art.","headline":"REAPS is a plausible incremental combination of CAM cropping and sequence learning, but its state-of-the-art claim does not survive a direct comparison with the omitted NTS-Net baseline.","tokens_in":9492,"tokens_out":1446,"would_cite":false,"duration_ms":17702,"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 single network, trained with only image-level labels, reaches state-of-the-art fine-grained recognition by attending to the object region and modeling its parts as a sequence.","keywords":["fine-grained image recognition","weakly supervised learning","class activation mapping","region attention","part sequence learning","bi-directional LSTM","soft parts","REAPS"],"falsifier":"A concrete check: on CUB-200-2011, compute the CAM bounding box at $\\tau=0.1$ for every test image and measure how often ground-truth part keypoints such as head, wing, or tail fall outside the crop; then retrain the pipeline under several values of $\\tau$. If a substantial fraction of diagnostic parts fall outside at $\\tau=0.1$, or if accuracy increases when the crop is enlarged, the central premise that the attended region preserves all discriminative details is disproved.","tokens_in":8464,"feed_emoji":"🖼️","tokens_out":6683,"duration_ms":63777,"temperature":0.7,"pith_summary":"REAPS is a unified network for fine-grained image recognition that needs only image-level category labels. It first uses class activation mapping to locate and amplify the object region, suppressing background, and then turns the region's feature map into a sequence of local descriptors that a bi-directional LSTM maps to implicit 'soft parts,' capturing spatial relationships among parts without fixing their number. Trained end-to-end on three benchmarks, it reports state-of-the-art accuracies of 93.5% on Stanford Cars, 92.6% on FGVC-Aircraft, and 86.8% on CUB-200-2011, beating methods that use bounding boxes or part annotations. The paper's claim is that part-based fine-grained recognition can be done more simply and more flexibly by replacing explicit part detectors with sequence modeling on an attended crop.","feed_headline":"Part-sequence model on an attended crop tops fine-grained benchmarks","feed_subtitle":"A single network, trained with only image-level labels, beats methods that need bounding boxes or part annotations on three datasets.","key_machinery":"The load-bearing mechanism is the 'soft-part' sequence model. After the CAM-based region attending network crops and zooms the attended region, the backbone's feature map $X\\in\\mathbb{R}^{H\\times W\\times C}$ is decomposed by a pooling operation into a sequence $Y=[Y_1,\\dots,Y_N]$ of $C$-dimensional vectors, each describing a rectangular region of the raw image. A bi-directional LSTM then maps this sequence into $M$ implicit part representations, and the concatenated hidden states serve as the part-level feature; this replaces hard part localization with a learned ordering over spatial descriptors. The second component, the class activation map from Eq. (1), is what suppresses background and defines the attended region that the sequence model sees.","core_discovery":"The central discovery is that two previously separate ideas, region-level attention and part-level modeling, can be fused into one end-to-end, weakly supervised pipeline that matches or exceeds stronger-supervised methods. On Stanford Cars, FGVC-Aircraft, and CUB-200-2011, the proposed REAPS+ reaches 93.5%, 92.6%, and 86.8% respectively. The largest margin over the prior state of the art appears on FGVC-Aircraft, where REAPS+ surpasses MA-CNN by a relative 2.7% gain; the authors attribute this to the sequence model exploiting the strong spatial and sequential structure of aircraft. The paper also reports that adding the part-sequence branch to a plain backbone gives a relative performance gain of about 1.0% on Stanford Cars, and that the region-attending component outperforms FCAN and RA-CNN in the single-scale comparison.","pith_inferences":["My inference: a testable extension the paper leaves open is to vary the CAM crop threshold $\\tau$ and measure accuracy; if performance is sensitive to $\\tau$, the method's robustness claim depends on a single unablated hyperparameter.","My inference: the same sequence mapping could be applied to other fine-grained domains where parts are naturally ordered, such as text lines, handwritten characters, or fingerprint ridges.","My inference: the paper's results on FGVC-Aircraft suggest, though the authors do not test it explicitly, that the sequence model's benefit grows with the regularity of part order, so categories with highly variable part arrangements might show smaller gains."],"forward_implications":["If REAPS is right, fine-grained recognition no longer needs bounding boxes or part keypoints at any stage; image-level labels are enough, which lowers annotation cost dramatically.","Because parts are implicit and their count is not fixed, the method should tolerate occlusion and incomplete objects better than approaches that detect a preset number of semantic parts.","Modeling parts as an ordered sequence lets the network exploit spatial context among parts, so categories with strong geometric ordering, such as aircraft, cars, and birds, stand to benefit most.","The framework is stackable: feeding the attended region from one PSN into a second PSN yields further gains of 93.5%, 92.6%, and 86.8%, suggesting that accuracy improves with depth of region refinement."],"supporting_citations":[{"why":"It supplies the class activation mapping that produces the region attention in the region attending network.","marker":"[40]"},{"why":"It is the recursive attention baseline whose single-scale accuracy is the direct comparison for the region attending network's effectiveness.","marker":"[9]"},{"why":"It is the prior multi-attention state of the art that REAPS+ surpasses on all three datasets.","marker":"[39]"},{"why":"It provides the LSTM used as the mapping function that turns serialized features into implicit part representations.","marker":"[13]"},{"why":"It is an attention localization baseline compared in the region attending network evaluation.","marker":"[23]"},{"why":"It supplies the Stanford Cars benchmark, one of the three evaluation datasets.","marker":"[18]"},{"why":"It supplies the FGVC-Aircraft benchmark, where the largest reported gains appear.","marker":"[24]"},{"why":"It supplies the CUB-200-2011 bird benchmark used for evaluation.","marker":"[29]"}],"fun_headline_variants":["Attention + part sequencing: one weak-supervised net beats fully-supervised","Fusing attention and part sequencing sets new standards in fine-grained","REAPS: end-to-end attention+sequencing takes three datasets","Weakly supervised attention+sequencing matches full supervision","Part-sequence model on attended crops tops three benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's accuracy rests on the assumption that the CAM-based attended region, cropped with threshold $\\tau=0.1$, contains every detail needed to tell the classes apart; if the crop leaves out a diagnostic part, the part-sequence branch never gets to see it.","fun_headline_variants_meta":{"raw":{"variants":["Attention + part sequencing: one weak-supervised net beats fully-supervised","Fusing attention and part sequencing sets new standards in fine-grained","REAPS: end-to-end attention+sequencing takes three datasets","Weakly supervised attention+sequencing matches full supervision","Part-sequence model on attended crops tops three benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000852,"raw_usage":{"total_tokens":3720,"prompt_tokens":981,"completion_tokens":2739,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":2655}},"tokens_in":597,"tokens_out":2739,"duration_ms":20346,"temperature":1.0,"reasoning_tokens":2655,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:58:06.592319+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete check: on CUB-200-2011, compute the CAM bounding box at $\\tau=0.1$ for every test image and measure how often ground-truth part keypoints such as head, wing, or tail fall outside the crop; then retrain the pipeline under several values of $\\tau$. If a substantial fraction of diagnostic parts fall outside at $\\tau=0.1$, or if accuracy increases when the crop is enlarged, the central premise that the attended region preserves all discriminative details is disproved.","supporting_citations":[{"cited_title":"In: CVPR","cited_arxiv_id":null,"evidence_quote":"It supplies the class activation mapping that produces the region attention in the region attending network."},{"cited_title":"In: CVPR","cited_arxiv_id":null,"evidence_quote":"It is the recursive attention baseline whose single-scale accuracy is the direct comparison for the region attending network's effectiveness."},{"cited_title":"In: ICCV","cited_arxiv_id":null,"evidence_quote":"It is the prior multi-attention state of the art that REAPS+ surpasses on all three datasets."},{"cited_title":"Neural Computation 9(8), 1735–1780 (1997)","cited_arxiv_id":null,"evidence_quote":"It provides the LSTM used as the mapping function that turns serialized features into implicit part representations."},{"cited_title":"In: CVPR","cited_arxiv_id":null,"evidence_quote":"It supplies the Stanford Cars benchmark, one of the three evaluation datasets."},{"cited_title":"HAL - INRIA (2013)","cited_arxiv_id":null,"evidence_quote":"It supplies the FGVC-Aircraft benchmark, where the largest reported gains appear."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the CUB-200-2011 bird benchmark used for evaluation."}],"review_version":1}