{"id":"68ecdad3-cf6b-4d12-9bab-3e59eeab39c7","arxiv_id":"2501.02640","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SAMPD combines adaptive modality weighting, contrastive pseudo-label enhancement, and retrieval-based augmentation to improve multispectral pedestrian detection under sparse annotations.","lead":"The paper proposes a teacher-student framework for detecting pedestrians in visible and thermal images when many ground-truth boxes are missing, using adaptive per-modality weights, a contrastive pseudo-label enhancement loss, and a retrieval module that pastes similar pedestrian patches into low-saliency image regions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The benchmark itself may be the confound: Algorithm 1's deterministic smallest-box removal, not real missing-label structure, could drive SAMPD's margins.","rationale":"The reader identified the synthetic sparse-annotation protocol as the weakest assumption. I agree, and the supplement sharpens the concern: Algorithm 1 uses `ARGMAX(PV)`, making the deletion deterministic rather than probabilistic, so the benchmark is not merely 'biased to small boxes' but is an extreme truncated-size distribution. The paper's own defense (supplementary Table 5) compares random removal only on the Supervised baseline, not on SAMPD or the strongest SAOD baselines, so it does not resolve the question of whether the reported margins are specific to the size-based deletion mechanism. The fully-annotated (0%) improvements are genuinely supportive evidence that the modules have some general value, but they do not directly test the sparse-annotation claim. The lack of error bars and the selection of PPE thresholds on the evaluation benchmark are secondary concerns; they could inflate or obscure the reported margins, but the more fundamental issue is whether the benchmark's missingness mechanism is representative. The reader's conditional verdict is appropriate, and this stress-test does not change it; it just adds a concrete way to verify external validity.","tokens_in":18419,"tokens_out":4416,"duration_ms":48729,"concrete_test":"Run the 30% scenario on KAIST under (i) random annotation deletion and (ii) deletion weighted by occlusion/visibility (e.g., using KAIST's occlusion attributes) instead of area, while keeping all methods and hyperparameters fixed. Evaluate SAMPD, Calibrated Teacher, and Supervised on the same test set with at least 3 random seeds. If SAMPD's MR margin over Calibrated Teacher in Table 1 (8.56 vs 10.47) shrinks to within seed noise or reverses under either alternative, the headline claim is protocol-dependent. Also report the resulting MR values and per-seed standard deviations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the sparse-annotation benchmark being representative of real missing labels. The supplementary Algorithm 1 defines that benchmark, and as written it is not a stochastic size-weighted deletion: `i <- ARGMAX(PV)` with `PV` inversely related to area means the smallest remaining box in each image is deterministically removed until the target count is reached. The result is a training set with a truncated box-size distribution, not a realistic pattern of annotator omissions. Because SAMPD's APRA module is specifically designed to re-introduce pedestrian patches and to convert high-confidence pseudo-labels into GT, it may be disproportionately well-matched to this 'all small boxes missing' protocol. The supplementary random-vs-ours comparison (Table 5) only measures the Supervised baseline, not SAMPD or Calibrated Teacher, so it does not establish that the claimed margins survive under a different missingness mechanism. The 0%-removal gains in Tables 5-6 are encouraging, but they do not directly validate the 30/50/70% sparse-setting claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SAMPD, a teacher-student framework for multispectral pedestrian detection under sparse annotation. It introduces three modules: MPAW, which weights per-modality detection losses by the cosine similarity between pseudo-label and ground-truth features; PPE, a contrastive loss that pulls high-quality pseudo-labels together and pushes them away from low-quality ones; and APRA, which augments training images with retrieved pedestrian patches and dynamically converts high-confidence pseudo-labels into ground-truth labels. The method is evaluated on KAIST and LLVIP with synthetically removed annotations (30%, 50%, 70%), reporting consistent MR reductions over prior SAOD methods, plus ablations and fully-annotated (0% removal) results.","tokens_in":18657,"tokens_out":3555,"duration_ms":36201,"significance":"If the reported results hold, SAMPD would be a useful contribution to sparsely annotated multispectral pedestrian detection, a practical setting where annotation is costly and incomplete. The paper's strengths include its clearly motivated modules, code release, consistent gains across two datasets and three removal ratios, and additional generalization experiments with different backbones and advanced MPD models (supplementary Tables 2 and 8). The fully-annotated results (Table 6) are also a nice bonus, showing the method can help even when labels are nominally complete. However, the central claim rests entirely on a synthetic annotation-removal protocol whose realism is not established, and the evaluation lacks statistical replication, which tempers the significance.","major_comments":[{"comment":"The sparse-annotation benchmark is the load-bearing element of the paper, but as written it is a deterministic procedure: Algorithm 1 repeatedly removes the single smallest remaining box in each image via ARGMAX(PV), not a stochastic size-weighted deletion. The main text describes 'increasing the probability of removing bounding-box annotations with smaller widths,' which contradicts the deterministic algorithm. More importantly, the supplementary comparison between random and size-based removal (Supplementary Table 5) is run only for the Supervised baseline, not for SAMPD or Calibrated Teacher, so it does not establish that the reported margins over baselines survive under a different missingness mechanism. Since APRA explicitly re-introduces pedestrian patches and MPAW/PPE re-weight pseudo-labels, the method may be disproportionately well-matched to the 'all small boxes missing' protocol. I request experiments under random removal for all compared methods, or a real-world sparsely annotated dataset, to support the general claim of improved performance in sparsely annotated environments.","section":"Supplementary Algorithm 1; Supplementary Table 5"},{"comment":"All reported numbers appear to come from a single training run, with no error bars, multiple seeds, or statistical tests. The improvements over the best prior method are modest in several cells (e.g., 8.56 vs. 10.47 MR at 30% removal on KAIST; 7.65 vs. 9.41 on LLVIP), and without variance estimates it is unclear whether these differences are meaningful. The abstract's word 'significantly' is not supported by any statistical evidence. Please report means and standard deviations over at least three seeds, or justify why single-run comparisons are standard for this task.","section":"Tables 1 and 2; Implementation Details"},{"comment":"The PPE thresholds τ1 = 0.9 and τ2 = 0.7 were selected by searching over several values and choosing the one that yields the best MR on the test set. This is a form of test-set overfitting, and it creates an unfair comparison with baselines whose hyperparameters were not tuned on this benchmark. The authors should either split a validation set for threshold selection or acknowledge that the reported gains may be optimistic. The threshold search also does not include error bars for each configuration, so the selection could be driven by noise.","section":"Supplementary Table 6; PPE module thresholds"},{"comment":"The description of the PPE loss is ambiguous about which model is updated. The text says the PG loss 'helps train the teacher model to generate higher-quality pseudo-labels,' but the loss uses feature vectors explicitly labeled as coming from the student model (l_{PL}^{(s)}, l_{GT}^{(s)}). If the teacher is an exponential moving average of the student, this should be stated; if the teacher has its own parameters, the gradient path for L_PG is unclear given Eq. (8) sums it into the total loss. This ambiguity affects reproducibility of the core method.","section":"Eq. (7) and surrounding text"}],"minor_comments":[{"comment":"There is a typo: 'approches' should be 'approaches.'","section":"Abstract"},{"comment":"The sentence 'the multispectral pedestrian-aware adaptive weight WAL in MPAW module' appears to contain a typo: 'WAL' is not defined anywhere; it likely should be 'wk' or 'MPAW.'","section":"Ablation Study (main text, after Table 3)"},{"comment":"The sentence 'Lsum_det denotes the detection loss using.' is incomplete; it should probably say 'using the MPAW-weighted combination in Eq. (3).'","section":"Eq. (8)"},{"comment":"The number of random seeds, validation split usage, and exact data augmentation details are not reported; this makes it harder to assess the stability and reproducibility of the results.","section":"Implementation Details"}],"recommendation":"major_revision","confidential_remarks":"The paper is potentially acceptable, but the central claim is currently supported only under a synthetic benchmark whose design is questionable and not tested against alternative missingness mechanisms. Adding random-removal experiments with all methods, error bars, and a fairer threshold-selection procedure would substantially strengthen the submission. The ambiguity about the teacher-student training dynamics in the PPE loss should also be resolved. If these issues are addressed, the paper could become a solid contribution to the SAOD and multispectral pedestrian detection communities."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Know this paper as a well-built engineering contribution to a niche problem, with one methodological wart that keeps its central claim from being fully trusted.\n\nWhat's new: the combination of three modules — adaptive per-modality pseudo-label weighting (MPAW), positive pseudo-label contrastive enhancement (PPE), and retrieval-based pedestrian augmentation with dynamic pseudo-label-to-GT conversion (APRA) — applied to multispectral pedestrian detection. Each component draws on existing ideas, but the specific assembly, especially APRA's brightness-matched patch retrieval for visible-thermal pairs, is not something I've seen before. The experiments are extensive: two datasets (KAIST, LLVIP), three removal ratios (30/50/70%), ablations for each module, plus fully-labeled 0% experiments and generalization to other backbones and detectors (MLPD, ProbEn). Results are consistently positive, and the 0% improvement (e.g., 7.58 to 6.50 MR on KAIST) is genuinely interesting, suggesting the method helps even when labels are complete. Code is released.\n\nThe load-bearing problem is the synthetic sparse-annotation protocol. Supplementary Algorithm 1 isn't a stochastic size-weighted deletion; it deterministically removes the smallest remaining box in each image until the target count is reached. That produces a training set with a truncated box-size distribution, which may not match how real annotators miss pedestrians (occlusion, low contrast, etc.). Because APRA is designed to re-introduce pedestrian patches and convert confident pseudo-labels to GT, it could be disproportionately well-matched to an 'all small boxes missing' benchmark. The supplementary random-vs-ours comparison (Table 5) only tests the Supervised baseline, not SAMPD or Calibrated Teacher, so it doesn't show the claimed margins survive under a different missingness mechanism. This is not a minor issue; it defines the benchmark on which every comparison is made.\n\nAlso, the PPE thresholds (tau1/tau2 = 0.9/0.7) were selected via a search over the evaluation set (supp Table 6), which inflates the headline numbers. And all results are single-run with no error bars or statistical tests. The latter is common in this subfield, but the former is a real overfitting risk.\n\nThat said, the equations are correct, the ablations are clean, and the fully-labeled gains are evidence that the method has value independent of the sparse protocol. The paper is honest enough to include a limitations paragraph and to report the threshold search.\n\nThis paper is for researchers working on SAOD or multispectral pedestrian detection. It deserves a serious referee, but the referee should treat the sparse-annotation simulation as the central point to probe. I'd want to see the method tested under random removal and under a more realistic annotation-noise model, plus repeated runs and confidence intervals.\n\nSend it for peer review, but be clear that the benchmark question needs a substantive answer before acceptance.","headline":"A well-engineered SAOD framework for multispectral pedestrian detection whose claimed gains rely on a synthetic deterministic small-box-removal benchmark that deserves more scrutiny.","tokens_in":19130,"tokens_out":3569,"would_cite":false,"duration_ms":30801,"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":"SAMPD claims that multispectral similarity between pseudo-labels and ground-truth boxes, plus adaptive retrieval augmentation, lets pedestrian detectors match nearly fully-supervised accuracy even when 30% of labels are removed.","keywords":["multispectral pedestrian detection","sparsely annotated object detection","pseudo-label quality","teacher-student learning","visible-thermal fusion","retrieval augmentation","miss rate"],"falsifier":"Re-run the same comparisons on a dataset with naturally sparse labels, or on KAIST/LLVIP with annotations removed according to occlusion or random patterns; if the MR advantage over Calibrated Teacher narrows or reverses (for example, if random removal makes SAMPD's 8.56 vs 10.47 gap shrink below measurement noise), the size-biased protocol is doing the work.","tokens_in":18250,"feed_emoji":"🚶","tokens_out":5369,"duration_ms":49337,"temperature":0.7,"pith_summary":"Multispectral pedestrian detectors assume every pedestrian is labeled, but real annotations are often incomplete, especially for small or occluded people. This paper argues that the two modalities themselves—visible texture plus thermal heat signature—can be used to tell trustworthy pseudo-labels from misleading ones, and that feeding the detector additional pedestrian patches matched to the scene can widen what it learns. The proposed framework, SAMPD, weights each modality's training loss by how similar student-model pseudo-label features are to ground-truth features, partitions pseudo-labels into positive, negative, and uncertain groups, and dynamically promotes high-quality pseudo-labels to ground-truth status. On the KAIST and LLVIP benchmarks with 30–70% of annotations artificially removed, it reports lower miss rates than prior sparsely-annotated detection methods, and it also improves the fully-supervised baseline, suggesting that even 'complete' labels contain gaps.","feed_headline":"Sparse labels no longer cripple multispectral pedestrian detection","feed_subtitle":"A visible-thermal teacher judges pseudo-label quality and cuts KAIST miss rate from 10.47 to 8.56 at 30% removal.","key_machinery":"The load-bearing object is the per-modality pseudo-label quality score $w_k = \\frac{1}{N}\\sum_{i=1}^{N}\\max_{j} d(l^{k(s)}_{PL_i}, l^{k(s)}_{GT_j})$, where $d$ is cosine similarity between global-average-pooled features of student-model pseudo-label boxes and ground-truth boxes, computed separately for visible, thermal, and fusion paths. This score both reweights detection losses and, through thresholds $\\tau_1 = 0.9$ and $\\tau_2 = 0.7$, sorts pseudo-labels into positive, negative, and uncertain categories for the positive pseudo-label guiding loss. The second mechanism is APRA, which retrieves pedestrian patches whose brightness matches the input image, pastes them into low-saliency regions where pedestrians are plausible, and dynamically converts high-quality pseudo-labels into ground truth. Together these turn pseudo-labels from noise into a curriculum for the student detector.","core_discovery":"On its own terms, the paper claims that sparse annotation is not a data-quantity problem but a label-quality problem, and that multispectral cues can repair it. SAMPD uses a teacher–student setup in which the teacher produces pseudo-labels for missing pedestrians; the student, trained on visible, thermal, and fused features, computes for each pseudo-label a cosine similarity to the nearest ground-truth feature. That single similarity score drives three mechanisms: it weights the per-modality detection loss (MPAW), it separates pseudo-labels into positive, negative, and uncertain groups for a guiding contrastive loss (PPE), and, when high, it promotes the pseudo-label into ground truth (APRA). The reported outcome is that at 30% removal SAMPD reaches 8.56 MR on KAIST and 7.65 on LLVIP, beating the best prior SAOD method by roughly 2 MR points, and at 0% removal it improves the fully labeled baseline from 7.58 to 6.50 MR on KAIST.","pith_inferences":["Because the synthetic protocol deletes small boxes preferentially, the reported margins likely overstate the benefit if real annotators miss pedestrians mainly because of occlusion, motion blur, or low contrast; testing on occlusion-sorted removal would separate those causes.","The cosine-similarity quality score is a generic idea: any detector with a feature backbone could weight pseudo-labels by their feature proximity to labeled objects, so the MPAW/PPE mechanism may transfer to single-modality SAOD tasks, not just multispectral pedestrians.","The paper's own limitation section notes APRA operates at image level using global brightness; a feature-level retrieval that matches individual pedestrian appearance rather than scene brightness could further reduce domain artifacts, and could be tested by replacing the brightness query with a per-patch appearance embedding."],"forward_implications":["At 30% label removal, SAMPD reports near-fully-supervised accuracy (8.56 MR on KAIST against 7.58 with all labels), meaning sparse annotations may cost little if pseudo-labels are quality-filtered by modality similarity.","The dynamic ground-truth refinement grows the annotation set during training (e.g., from 40,107 to 47,369 boxes on KAIST), closing gaps that otherwise make the visible and thermal streams disagree.","The framework transfers to other backbones and detectors: supplementary results show gains with ResNet backbones and with MLPD and ProbEn as base detectors, not just the SSD+VGG16 used in the main tables.","The method also improves the fully annotated setting (KAIST MR from 7.58 to 6.50; LLVIP from 6.01 to 5.70), which the paper interprets as evidence that real-world 'complete' labels still contain missing pedestrians."],"supporting_citations":[{"why":"Supplies the KAIST visible-thermal pedestrian dataset that anchors the main comparisons and the fully-supervised baseline.","marker":"Hwang et al. 2015"},{"why":"Supplies the LLVIP low-light visible-infrared dataset used as the second benchmark.","marker":"Jia et al. 2021"},{"why":"Introduces the pseudo-label approach to sparsely annotated detection and the removal-ratio convention (30/50/70%) that SAMPD follows.","marker":"Niitani et al. 2019"},{"why":"Calibrated Teacher is the strongest prior SAOD baseline; SAMPD must beat its calibrator-based pseudo-label validation.","marker":"Wang et al. 2023a"},{"why":"SparseDet supplies the size-biased sparse-annotation simulation protocol and a self-supervised-loss baseline that SAMPD extends.","marker":"Suri et al. 2023"},{"why":"Robust Teacher is the pedestrian augmentation baseline that APRA is compared against in the augmentation ablation.","marker":"Li et al. 2023"},{"why":"MLPD provides the multi-label pedestrian detection loss and the SSD+VGG16 architecture used as the foundation of SAMPD.","marker":"Kim et al. 2021"},{"why":"SSD is the single-shot detector architecture that the teacher and student models are built on.","marker":"Liu et al. 2016"}],"fun_headline_variants":["Teacher-student model turns sparse labels into rich pedestrian data","Sparse annotations? Multispectral teacher fixes them","SAMPD: quality pseudo-labels for sparse multispectral detection","Visible-thermal cues rescue pedestrian detection from sparse labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole evaluation runs on a synthetic protocol that removes small bounding boxes with higher probability than large ones, so the reported gains assume real annotation gaps follow that size-based pattern rather than, say, random or occlusion-driven omissions.","fun_headline_variants_meta":{"raw":{"variants":["Teacher-student model turns sparse labels into rich pedestrian data","Sparse annotations? Multispectral teacher fixes them","SAMPD: quality pseudo-labels for sparse multispectral detection","Visible-thermal cues rescue pedestrian detection from sparse labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000188,"raw_usage":{"total_tokens":1358,"prompt_tokens":995,"completion_tokens":363,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":611,"completion_tokens_details":{"reasoning_tokens":297}},"tokens_in":611,"tokens_out":363,"duration_ms":4132,"temperature":1.0,"reasoning_tokens":297,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:07:03.438046+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same comparisons on a dataset with naturally sparse labels, or on KAIST/LLVIP with annotations removed according to occlusion or random patterns; if the MR advantage over Calibrated Teacher narrows or reverses (for example, if random removal makes SAMPD's 8.56 vs 10.47 gap shrink below measurement noise), the size-biased protocol is doing the work.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the KAIST visible-thermal pedestrian dataset that anchors the main comparisons and the fully-supervised baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the LLVIP low-light visible-infrared dataset used as the second benchmark."},{"cited_title":"Sampling Techniques for Large-Scale Object Detection from Sparsely Annotated Objects","cited_arxiv_id":"1811.10862","evidence_quote":"Introduces the pseudo-label approach to sparsely annotated detection and the removal-ratio convention (30/50/70%) that SAMPD follows."},{"cited_title":"SparseDet: Improving Sparsely Annotated Object Detection with Pseudo-positive Mining","cited_arxiv_id":"2201.04620","evidence_quote":"SparseDet supplies the size-biased sparse-annotation simulation protocol and a self-supervised-loss baseline that SAMPD extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SSD is the single-shot detector architecture that the teacher and student models are built on."}],"review_version":1}