{"id":"0e50f82e-8c4a-48eb-9060-de30e7885b31","arxiv_id":"2412.00818","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A diffusion-based keypoint propagation plus categorical keypoint positional embedding is reported to improve animal ReID accuracy, but the evidence is limited to a single baseline comparison.","lead":"This paper proposes using GPT-4 to label three keypoints on one animal photo, a pre-trained diffusion model to propagate those keypoints to all photos, and a Vision Transformer with keypoint and categorical keypoint embeddings for animal re-identification. The authors report large accuracy gains on four wildlife datasets but compare only against one baseline, leaving the central performance claim under-supported.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central outperformance claim rests on unverified accuracy numbers: only one toolkit baseline is compared, and the retrieval metric is defined via a binary-classification formula with no top-1 rule or error bars.","rationale":"The reader's verdict is REJECT, and I agree with that conclusion. My concern is more precise than the stated weakest assumption. The keypoint-propagation localization issue is a genuine internal-validity concern, because Section 3.2's core mechanism is never directly evaluated. However, it is secondary to the empirical claim: end-to-end accuracy could in principle remain high even if propagated keypoints are noisy, especially if the ViT's own features dominate, and the random-point ablation (Tables 6 and 7) only partially addresses this. The evaluation protocol is more load-bearing because a wrong or non-comparable metric invalidates every reported number, including the comparison that motivates the paper. The paper itself flags the ambiguity by stating a binary-classification accuracy formula for a retrieval task, and Table 1 contains a single baseline row. A single computational check—recomputing Rank-1 under a standard protocol and comparing to a published baseline—would settle whether the reported gains are real. Thus I recommend no change to the reader's REJECT verdict.","tokens_in":12703,"tokens_out":5426,"duration_ms":50114,"concrete_test":"Using the released code or a reimplementation of Section 3.5, run the same 80/20 identity split and compute both (a) the paper's TP/TN/FP/FN accuracy and (b) the standard Rank-1 accuracy from the WildlifeDatasets/PetFace protocols on all four datasets. Then compare against at least one published baseline from [34] (e.g., MegaDescriptor) under identical query/gallery settings. If (a) differs materially from (b), or if a published baseline matches or exceeds the CKPE row in Table 1 on any dataset, the central outperformance claim is not substantiated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: CKPE 'significantly outperforms existing state-of-the-art methods' with accuracy gains from +5.9% to +50.1% (Abstract, Section 4.2). For that claim to hold, the numbers in Table 1 must be correct and must be measured on the same metric as the SOTA they are compared against. Both conditions are unsecured. (1) The only comparison row is 'WildLife Dataset[34]', which is an open-source toolkit and baseline, not a published SOTA method; no method from the Section 2 related work is included, so the 'previous SOTA' values (e.g., 41.0% on Giraffe) are not established as SOTA. (2) Section 4.1 defines Accuracy = (TP+TN)/(TP+TN+FP+FN), a binary-classification formula, but Section 3.5 describes a retrieval protocol (cosine similarity ranking against a gallery). The paper never specifies how the gallery's many negative pairs are counted as TN or how the top-1 decision is converted into TP/TN/FP/FN. Under a naive pair-counting interpretation, accuracy is dominated by true negatives and can be high even with poor ranking. (3) No error bars or repeated runs are reported, so the 91.1% versus 41.0% Giraffe jump is a single unverified point. Because the entire claim is about outperforming SOTA, the evaluation protocol is the load-bearing element.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a pipeline for animal re-identification in which GPT-4 selects a small set of discriminative keypoints on a single image, a pre-trained Stable Diffusion model propagates those keypoints across the dataset, and the resulting keypoint locations are injected into a Vision Transformer through Keypoint Positional Embedding (KPE) or Categorical Keypoint Positional Embedding (CKPE). The authors report large accuracy gains over a single toolkit baseline on MacaqueFaces, Giraffe, IPandas, and CTai, and they present ablations on keypoint source, diffusion timestep, ViT scale, keypoint count, and random-point controls.","tokens_in":12971,"tokens_out":5717,"duration_ms":50542,"significance":"If the reported results are reliable, the proposed low-annotation keypoint propagation mechanism would be a practically valuable contribution to wildlife ReID, where manual keypoint annotation is expensive. The idea of using diffusion-feature correspondence for keypoint propagation is timely, and the KPE/CKPE embedding is simple and interpretable. However, the manuscript currently does not establish its central claim: the evaluation is limited to a single toolkit baseline, the accuracy metric is not defined in a way consistent with the retrieval protocol, and the ablation tables contain internal inconsistencies. The paper contains no machine-checked proofs or released code, so the contribution rests entirely on the unvalidated experimental numbers.","major_comments":[{"comment":"The evaluation metric is not consistent with the testing protocol. Section 4.1 defines Accuracy = (TP+TN)/(TP+TN+FP+FN), which is a binary-classification formula, but Section 3.5 describes a retrieval protocol in which the gallery image with highest cosine similarity to the query is selected. The manuscript never specifies how TP, TN, FP, and FN are counted for a multi-gallery retrieval task, nor how the top-1 decision is converted into these counts. Under a naive pair-counting interpretation the accuracy would be dominated by true negatives and could be high even with poor ranking. The numbers in Table 1 are therefore uninterpretable until the protocol is specified unambiguously; the authors should report Rank-1, Rank-5, or mAP, or explicitly define the binary decision rule.","section":"§4.1"},{"comment":"The central claim of outperforming state-of-the-art methods is not supported by the comparison. The only baseline row is 'WildLife Dataset[34]', which is an open-source toolkit baseline, not a published SOTA method. The related work cites specialist methods (e.g., ATRW [17], Miele et al. [22]) and recent general approaches (e.g., 'Toward re-identifying any animal' [12], PetFace [31]), but none of these are included in Table 1. Consequently, the values described as 'previous SOTA' (e.g., 41.0% on Giraffe) are not established as SOTA, and the claimed improvements of +5.9% to +50.1% are not calibrated against the methods the paper itself identifies as relevant prior work. New experiments with published SOTA methods under an identical, clearly defined protocol are required.","section":"§4.2, Table 1"},{"comment":"The ablation tables contain an internal inconsistency that undermines the keypoint-ablation claims. Table 4 reports that ViT-Base without keypoints achieves 92.5%, while Table 5 reports 97.3% for the 'NO keypoint' condition on the same dataset and presumably the same settings. If these are different experimental configurations, that must be stated; if they are intended to be the same, one of the tables is wrong. This discrepancy matters because the paper uses these numbers to argue that keypoints improve performance, and a 4.8-point difference in the baseline changes the size of the claimed improvement. In addition, no error bars or repeated runs are reported anywhere, so differences of 0.3–0.6 points in Tables 2–7 cannot be distinguished from noise.","section":"§5.3 and §5.4, Tables 4 and 5"},{"comment":"The categorical embedding equation is dimensionally unclear. The paper defines CMI = OneHot(PI) ∈ R^{(K+1)×Nc} and then states CMI · Wckp ∈ R^d. If CMI is a matrix of shape (K+1) × Nc and Wckp is Nc × d, the product is a matrix of shape (K+1) × d, not a vector in R^d. If the intended operation is a sum or mean over keypoints or patches, that operation is not specified. The same issue affects the KPE description in Section 3.3, where multiple keypoints would produce multiple mask vectors but the aggregation into a single additive embedding is not defined. The architecture is not fully specified without this detail.","section":"§3.4"},{"comment":"The keypoint propagation step, which is the foundation of the method, is never validated against keypoint ground truth. The paper claims that cosine similarity between diffusion features 'ensures consistent keypoint localization even when faced with varying poses, occlusions, and environmental conditions', but no keypoint localization accuracy, correspondence accuracy, or visual evaluation is reported. Since the KPE and CKPE embeddings depend entirely on the propagated keypoints, a failure of propagation would directly affect the ReID results. The authors should measure keypoint localization quality (e.g., PCK) or at least provide a quantitative analysis of propagation failures across the four datasets before the mechanism can be considered supported.","section":"§3.2"}],"minor_comments":[{"comment":"The claimed improvement range '+5.9% to +50.1%' does not match Table 1: the smallest gap in Table 1 is 6.2% (MacaqueFaces) and the largest is 50.1% (Giraffe). The range should be corrected or the source of the 5.9% figure should be explained.","section":"Abstract and §1"},{"comment":"The sentence 'aligning with similar findings reported in []' contains an empty citation; the reference is missing.","section":"§5.2"},{"comment":"The text states that both KPE and CKPE 'consistently improve performance across all ViT scales'. At ViT-Base, CKPE (97.3%) is lower than KPE (98.5%), so the wording 'consistently improve' should be qualified to describe the comparison against the no-keypoint baseline rather than between KPE and CKPE.","section":"§5.3, Table 4"},{"comment":"The table header 'NO keypoint 1 3 6 10' is ambiguous: it is not clear from the caption whether the first column is the no-keypoint baseline or whether the columns correspond to different numbers of keypoints for each ViT scale. The table should be relabeled.","section":"§5.4, Table 5"},{"comment":"The section title contains a typo: 'Propagration' should be 'Propagation'.","section":"§3.2"},{"comment":"The sentence 'we will made our codebase, datasets, and detailed annotations publicly available' is grammatically incorrect and should be rewritten. Also, the promised code/data releases are not currently available, so the reproducibility claim cannot be verified.","section":"§1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of a computer-vision venue, and the core idea is plausible, but the evaluation currently does not support the headline claims. The primary issues are the ill-defined accuracy metric, the absence of published SOTA baselines, and the internal inconsistency between Tables 4 and 5. These are fixable with additional experiments and a rewritten evaluation section, so I do not recommend rejection outright; however, the current numbers should not be cited or accepted until the protocol is clarified and the comparisons are expanded."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Briefly: the core idea is genuinely attractive — you annotate one image per species, propagate keypoints through a frozen diffusion model, and feed the locations plus categories to a ViT via KPE/CKPE. That could reduce annotation cost, which matters for ecological monitoring. The CKPE embedding itself is a real addition over the keypoint-promptable human ReID work they cite [32].\n\nWhere it earns credit: the ablations are sensible, especially comparing GPT-4 keypoints against expert annotations and checking timestep sensitivity. The writing is clear enough that you can see what they did. They also acknowledge the diffusion correspondence is borrowed from prior work.\n\nBut the headline numbers don't hold up. Table 1 compares against one row, 'WildLife Dataset' — a toolkit baseline, not the published methods they discuss in Section 2. No error bars anywhere. The accuracy formula in Section 4.1 is binary classification, yet the protocol is retrieval; they never say how true negatives are counted. Under any natural pair-counting interpretation, accuracy is dominated by negatives and the 91.1% vs 41.0% Giraffe jump becomes uninterpretable. That's the load-bearing flaw.\n\nThere are also smaller internal problems. The CMI equation gives a (K+1)×Nc matrix, but the text calls it an embedding in R^d; the multiplication with Wckp produces a matrix, not a vector, and the addition to patch features doesn't match. The keypoint-count ablation (Table 5) shows performance dropping as keypoints go from 3 to 10 on large/huge, which they explain away but doesn't support the argument that more semantic keypoints help. There's an empty citation in Section 5.2 and the tables are malformed in the preprint.\n\nThe paper isn't a waste of time. Read it for the annotation-reduction idea and the diffusion propagation details; those are worth knowing. But treat every reported accuracy as unverified until they rerun against real baselines, report variance, and correct the metric definition.\n\nMy recommendation: send it to peer review if the venue tolerates heavy revision — the idea deserves scrutiny — but expect that it comes back rejected unless the evaluation is redone. If I were the editor, I'd ask for major revision rather than desk-reject, because the flaws are fixable and the core concept is novel enough to be worth the effort.","headline":"Promising keypoint-propagation idea undermined by a comparison table with one baseline and an ambiguous accuracy metric.","tokens_in":13537,"tokens_out":3773,"would_cite":false,"duration_ms":31428,"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":"Propagated keypoints lift animal re-ID accuracy by up to 50.1 percent.","keywords":["animal re-identification","keypoint propagation","diffusion model correspondences","Vision Transformer","categorical keypoint positional embedding","wildlife datasets","semantic keypoints","GPT-4 keypoint selection"],"falsifier":"Annotate a held-out sample from the four datasets with expert keypoint locations and measure the distance between each diffusion-propagated keypoint and the true landmark; if a substantial fraction land more than one ViT patch width away, the correspondence assumption that carries the accuracy gains is violated.","tokens_in":1519,"feed_emoji":"🐾","tokens_out":3004,"duration_ms":68614,"temperature":0.7,"pith_summary":"This paper tries to show that animal re-identification—matching photographs of the same individual across camera traps—can be made both more accurate and far cheaper to annotate. It proposes a pipeline that uses one human- or LLM-annotated image per species, propagates those keypoints to every other image through a pre-trained diffusion model, and then injects the keypoint locations and their semantic categories into a Vision Transformer via trainable embeddings. Across four wildlife benchmarks the authors report accuracy gains over the previous state of the art ranging from +5.9 to +50.1 percentage points. If the result holds, ecologists could deploy species-agnostic ReID with minimal manual labeling.","feed_headline":"Propagated keypoints lift animal re-ID accuracy by up to 50.1%","feed_subtitle":"A single GPT-4-labeled image plus a diffusion model lets a ViT match individuals across four wildlife datasets.","key_machinery":"The load-bearing mechanism is keypoint propagation through emergent diffusion correspondence, combined with Categorical Keypoint Positional Embedding (CKPE). A single reference image's keypoints, selected by GPT-4, are located in every other image by taking the argmax of cosine similarity between the reference keypoint's diffusion feature vector and the target image's diffusion feature map. The resulting keypoint locations are encoded as one-hot masks over ViT patches in Keypoint Positional Embedding (KPE) and, per keypoint category, as a second one-hot axis in CKPE; learned matrices project each mask into the embedding dimension and the vectors are added to the patch features. These additions let the ViT attend to semantically labeled anatomical locations without any per-image keypoint supervision.","core_discovery":"The central claim is that keypoint information, obtained without per-image annotation, substantially improves animal re-identification when encoded as positional and categorical embeddings in a Vision Transformer. Keypoints are chosen once per dataset by GPT-4 and then propagated to all images by matching diffusion-model features: for each keypoint in the reference image, the location in a target image is the argmax of cosine similarity between the reference keypoint feature vector and the target feature map. These locations are turned into one-hot patch masks, embedded through learned matrices, and added to ViT patch features; the categorical variant uses a separate one-hot axis per keypoint type such as left eye, right eye, or nose. The trained model matches query images to gallery images by cosine similarity of the [CLS] token. The authors report accuracies of 99.6% on MacaqueFaces, 91.1% on Giraffe, 90.3% on IPandas, and 83.4% on CTai, exceeding the WildlifeDatasets baseline by 6.2, 50.1, 21.0, and 24.2 points respectively.","pith_inferences":["The propagation mechanism is not tied to GPT-4 or to Stable Diffusion specifically; any vision-language model that can name landmarks and any self-supervised correspondence model could plausibly replace them, which is a testable extension.","The same one-shot keypoint propagation plus categorical embedding recipe could transfer to other fine-grained instance-recognition tasks, such as identifying individual humans from non-frontal views or matching animals in camera-trap sequences.","The paper's causal story, that semantic keypoints rather than mere points drive the gains, would be strengthened by measuring keypoint localization accuracy directly; the current evidence is end-to-end accuracy only.","If diffusion features are what make propagation work, the choice of diffusion timestep matters; the authors find timesteps 0-300 work best, suggesting that coarse semantic features, not fine details, carry the correspondence."],"forward_implications":["If the reported gains hold, wildlife ReID can be deployed with only one manually annotated image per species, cutting annotation cost by orders of magnitude.","Keypoint categories such as left eye versus nose provide a semantic axis that helps the ViT separate individuals, and ablation results show the categorical variant consistently beats purely positional keypoints.","The method is species-agnostic: the same GPT-4-plus-diffusion keypoint pipeline, without retraining the propagation step, is applied to macaques, giraffes, pandas, and chimpanzees.","Because the keypoint embeddings operate on patch tokens, the approach scales with ViT size, with the largest gains on the smallest ViT-Base model."],"supporting_citations":[{"why":"supplies the emergent correspondence property of image diffusion models that the keypoint propagation relies on","marker":"[33]"},{"why":"demonstrates unsupervised semantic correspondence with Stable Diffusion, the specific mechanism used to match keypoints","marker":"[11]"},{"why":"introduces keypoint-promptable ReID with a transformer, the approach this paper adapts from humans to animals","marker":"[32]"},{"why":"provides the WildlifeDatasets baseline numbers and the experimental setting the paper compares against","marker":"[34]"},{"why":"defines the accuracy metric and evaluation protocol used for the reported comparisons","marker":"[30]"},{"why":"supplies the Giraffe dataset where the largest gain (41.0% to 91.1%) is reported","marker":"[22]"}],"fun_headline_variants":["Diffusion-propagated keypoints boost animal re-ID by up to 50.1%","One labeled image + diffusion model lifts animal re-ID by up to 50.1%","One-image keypoint embeddings lift animal re-ID by up to 50.1%","Diffusion model spreads keypoints from one image, lifting animal re-ID by up to 50.1%"],"cache_read_input_tokens":15616,"weakest_assumption_plain":"The pipeline's load-bearing premise is that diffusion-feature cosine similarity puts each propagated keypoint on the same anatomical landmark in every image, a premise the paper never tests directly since it reports only end-to-end ReID accuracy.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion-propagated keypoints boost animal re-ID by up to 50.1%","One labeled image + diffusion model lifts animal re-ID by up to 50.1%","One-image keypoint embeddings lift animal re-ID by up to 50.1%","Diffusion model spreads keypoints from one image, lifting animal re-ID by up to 50.1%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00151,"raw_usage":{"total_tokens":6069,"prompt_tokens":978,"completion_tokens":5091,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":594,"completion_tokens_details":{"reasoning_tokens":4992}},"tokens_in":594,"tokens_out":5091,"duration_ms":30096,"temperature":1.0,"reasoning_tokens":4992,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:58:35.142330+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Annotate a held-out sample from the four datasets with expert keypoint locations and measure the distance between each diffusion-propagated keypoint and the true landmark; if a substantial fraction land more than one ViT patch width away, the correspondence assumption that carries the accuracy gains is violated.","supporting_citations":[{"cited_title":"Emergent correspondence from image diffusion","cited_arxiv_id":null,"evidence_quote":"supplies the emergent correspondence property of image diffusion models that the keypoint propagation relies on"},{"cited_title":"Keypoint promptable re-identification","cited_arxiv_id":null,"evidence_quote":"introduces keypoint-promptable ReID with a transformer, the approach this paper adapts from humans to animals"},{"cited_title":"WildlifeDatasets: An Open-Source Toolkit for Animal Re-Identification","cited_arxiv_id":null,"evidence_quote":"provides the WildlifeDatasets baseline numbers and the experimental setting the paper compares against"},{"cited_title":"Revisiting animal photo-identification using deep metric learning and network analysis","cited_arxiv_id":null,"evidence_quote":"supplies the Giraffe dataset where the largest gain (41.0% to 91.1%) is reported"}],"review_version":1}