{"id":"73506305-2c4a-4606-ba84-c4dd410b6b73","arxiv_id":"2505.06937","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"TAPNet fuses RGB and thermal imagery using attention and feature-decomposition modules and reports improved crowd counting and localization on two UAV datasets.","lead":"This paper describes TAPNet, a network that fuses visible and infrared images to count and locate people in dense crowds filmed from drones, including at night. The authors report lower counting errors than several baselines on two RGB-thermal datasets, but the evaluation has internal inconsistencies and no independent test set.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2's final R-T row is internally inconsistent: the DroneRGBT MAE matches DAFP-only, its MSE/F1 match the Table 5 (1,0) hyperparameter setting, and the GAIIC2 row matches AFDF+Spatial Shift only. No table reports the complete DAFP+AFDF model.","rationale":"I agree with the reader's overall rejection, but the single most load-bearing concern is sharper than the reader's stated weakest assumption about spatial offset augmentation. The reader's rationale already noted Table 2/Table 3 discrepancies and the GAIIC2 result matching only the AFDF ablation; those internal inconsistencies are the real load-bearing issue because the central claim is an empirical comparison that requires the reported numbers to describe one complete architecture. The augmentation-distribution assumption would matter only after the headline numbers are established as belonging to the proposed model. Since Table 2's R-T rows are not reproducible from any consistent full-model configuration reported in Tables 3, 5, or 8, the comparison against P2PNet, CLTR, PET, APGCC, DroneNet, and MMCount is unsubstantiated as presented. The concrete test I propose is direct and minimal: run the four configuration variants using the released code and check whether the complete DAFP+AFDF model with spatial shift reproduces the Table 2 rows. This is a reproducibility check that does not depend on judging the novelty of the fusion modules or on whether the augmentation is theoretically motivated. I therefore keep the reader's REJECT verdict, with the primary justification being the internal metric inconsistency rather than the specific form of the random offset augmentation.","tokens_in":15842,"tokens_out":4271,"duration_ms":42624,"concrete_test":"Using the released repository (https://github.com/zz-zik/TAPNet), retrain or load checkpoints for four configurations under identical Section 5.3 hyperparameters: (1) DAFP only, (2) AFDF only, (3) DAFP+AFDF combined per Fig. 2, and (4) DAFP+AFDF with spatial random offset augmentation. Evaluate each on the DroneRGBT and GAIIC2 validation splits and compare against Table 2, Table 3, and Table 8. The concern is settled if configuration (4) reproduces the exact reported rows; in particular, DroneRGBT 7.32/11.54/0.657 and GAIIC2 7.87/13.25/0.526. If the reproduced full-model metrics differ by more than rounding precision in any of the three metrics, or if the GAIIC2 row is reproduced only by AFDF+Spatial Shift, then the headline comparison is not supported by the reported experiments.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is an empirical superiority claim for TAPNet, and its sole evidence is Table 2. For that claim to hold, the R-T row of Table 2 must describe the complete proposed architecture (DAFP+AFDF with spatial random offset augmentation) under the fixed hyperparameters of Section 5.3. The paper's own tables contradict this. On DroneRGBT, Table 2 reports MAE/MSE/F1 = 7.32/11.54/0.657. Section 6.2's Table 3 gives R-T+DAFP = 7.32/11.71/0.697 and R-T+AFDF = 7.51/12.06/0.712. Thus the Table 2 MAE equals the DAFP-only MAE, but its MSE and F1 are not the DAFP-only values; they exactly match the Table 5 row with auxiliary points (kp,kn) = (1,0), a hyperparameter search on the same validation set. On GAIIC2, Table 2's 7.87/13.25/0.526 is identical to the 'AFDF+Spatial Shift' row of Table 8, not to any row combining DAFP and AFDF. Together, these inconsistencies mean the headline numbers cannot be attributed to a single, well-defined model configuration. If the final model is AFDF-only with spatial shift, the claimed dual-module contribution is not demonstrated; if it is DAFP+AFDF, no reported ablation supports the final row. This is not a question of consensus or novelty; it is an internal traceability failure in the central quantitative claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TAPNet, a transformer-based RGB-T crowd counting and localization network that combines a dual-optical attention fusion module (DAFP), an adaptive feature decomposition fusion module (AFDF), and a spatial random offset data augmentation strategy. The authors claim state-of-the-art results on DroneRGBT and GAIIC2, with best RGB-T results of 7.32/11.54/0.657 and 7.87/13.25/0.526 in MAE/MSE/F1, respectively. The manuscript includes architecture details, loss formulations, and ablation studies on fusion modules, auxiliary point settings, and the offset augmentation.","tokens_in":16202,"tokens_out":4650,"duration_ms":42031,"significance":"If the reported results were reliable, the paper would make a useful contribution to multimodal crowd counting, particularly for low-light and misaligned RGB-TIR image pairs, and the public release of code would aid reproducibility. The topic is relevant, and the point-based localization with dual-modal fusion is a reasonable direction. However, the central empirical claim is undermined by internal inconsistencies across the tables and by a potentially circular evaluation protocol; the evidence as presented does not establish the stated superiority of the full TAPNet architecture.","major_comments":[{"comment":"The final R-T row in Table 2 is not traceable to a single well-defined model configuration. On DroneRGBT, Table 2 reports MAE/MSE/F1 = 7.32/11.54/0.657, but Table 3 reports R-T+DAFP = 7.32/11.71/0.697; the MSE and F1 in Table 2 exactly match the (kp,kn) = (1,0) row of Table 5, not the DAFP-only row. On GAIIC2, Table 2's 7.87/13.25/0.526 is identical to the 'AFDF+Spatial Shift' row of Table 8, not to any configuration combining DAFP and AFDF. Thus the headline superiority claim cannot be attributed to the complete TAPNet architecture, and no ablation in the paper reports the full model.","section":"Section 6, Tables 2, 3, 5, 8"},{"comment":"The spatial random offset augmentation is explicitly stated in Section 5.3 to 'enable the validation set to have the same image offset distribution as the training set.' Since the final metrics are reported on that same validation set, the improvement attributed to the augmentation may reflect fitting the validation distribution rather than a generalizable alignment mechanism. The paper provides no measurement of the actual RGB-TIR misalignment in either dataset and no evaluation on a held-out test set. This circularity is a load-bearing issue for the claimed generalization benefit.","section":"Sections 4.4 and 5.3"},{"comment":"Hyperparameters for auxiliary point counts (kp, kn) and random ranges (npos, nneg) are selected on the same validation set that is used for the final reported metrics, with no separate validation split. Combined with the internal inconsistency in Table 2, this makes it difficult to rule out overfitting to the benchmark as the source of the reported gains. A separate validation set or cross-validation should be used for hyperparameter selection.","section":"Section 5.3 and Tables 5–6"}],"minor_comments":[{"comment":"The equation lists two identical expressions for w'_{c2}; the first should presumably be w'_{c1}.","section":"Equation (12)"},{"comment":"The text refers to 'the spatially randomized offset data enhancement strategy proposed in 2.4'; this should reference Section 4.4.","section":"Section 5.3"},{"comment":"There is a typo: 'Transorfmer shared layer' should read 'Transformer shared layer'.","section":"Section 4.3"},{"comment":"The decomposition loss Lcc uses parameters α and β that are not defined in the surrounding text; please clarify whether these are the same as the modality fusion weights in Equation (13).","section":"Equation (25)"},{"comment":"The caption says 'Dual Fusion Module' but the table actually compares two separate modules, DAFP and AFDF, evaluated independently; the caption should be clarified.","section":"Table 3 caption"},{"comment":"Several references list only 'et al.' without author names (e.g., [1], [3], [4], [5], [7], [9], [10], [12], [14], [15], [16], [17], [18], [19], [20], [23], [24], [25], [27], [29], [30], [31], [32], [33], [34], [35], [37], [38], [39], [40], [41], [43], [44], [45]); please complete the author lists.","section":"References"}],"recommendation":"reject","confidential_remarks":"The internal inconsistencies in Table 2 suggest that the reported headline numbers may have been assembled from different experimental configurations. Even if the authors can correct the reporting, the circular use of the validation set for both augmentation design and final evaluation, together with the absence of any experiment on the combined DAFP+AFDF model, leaves the central empirical claim unsupported. A rejection with encouragement to resubmit after a thorough re-run of experiments with a proper evaluation protocol seems appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline numbers in Table 2 don't survive contact with the paper's own ablations. On GAIIC2, the final R-T row (7.87/13.25/0.526) exactly matches the AFDF+Spatial Shift row in Table 8, not any configuration with the proposed DAFP module. On DroneRGBT, the final row (7.32/11.54/0.657) exactly matches the (1,0) auxiliary-point setting in Table 5, but not the DAFP-only row in Table 3 (7.32/11.71/0.697). No table reports the full DAFP+AFDF model that the abstract and method sections describe. That is a load-bearing traceability failure, not a cosmetic inconsistency.\n\nThere is something real here, though. RGB-T crowd counting from UAV views with misaligned modality pairs is a legitimate and under-served problem. Adapting APGCC's point-based matching framework to RGB-T is a sensible direction, and the random-shift augmentation is a simple idea worth testing. The authors also provide a code link, which is more than many papers do.\n\nThe soft spots run deeper than the table mismatch. The spatial random offset augmentation is explicitly designed to match the validation set's offset distribution (Sections 4.4 and 5.3), and the final metrics are reported on that same validation set—so the reported gains are partly an artifact of tuning the augmentation to the benchmark. Hyperparameters are selected on the same validation split, with no error bars and no independent test set. The GAIIC2 ground truth is author-created and not released, so independent verification is impossible. There are also minor notation slips in the equations and some typos, but those are not the reason to be skeptical.\n\nWho is this for? Someone building a baseline in RGB-T crowd counting might look at the code and maybe reuse the augmentation idea. But the reported numbers should not be trusted until the paper defines a single model, re-runs the experiments on a proper train/test split, and includes an ablation that actually includes the full proposed architecture.\n\nMy recommendation: I would not send this to peer review as it stands. It needs a major revision with re-experiments first. If a revised version appears with clean, traceable results, it would be worth another look.","headline":"Table 2's headline R-T results don't match any single model in the paper's own ablations, so the central empirical claim is unsubstantiated.","tokens_in":16808,"tokens_out":5581,"would_cite":false,"duration_ms":52701,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that fusing infrared with visible light, using two attention-style fusion modules plus a spatial-offset augmentation, lowers crowd counting error on UAV RGB-TIR benchmarks, with MAE falling from 10.32 to 7.32 on DroneRGBT.","keywords":["crowd counting","RGB-thermal fusion","point-based localization","attention fusion","UAV imagery","misalignment augmentation","multi-kernel MMD","transformer"],"falsifier":"Measure the actual per-pair displacement between RGB and TIR in DroneRGBT and GAIIC2, for example by matching annotated head points or estimating homographies: if typical misalignment exceeds 10 pixels or varies across the image, the augmentation's assumed distribution is wrong, and the reported gains should not transfer to other misaligned datasets. Retraining with offsets drawn from a wider or adaptive distribution and comparing MAE on an independent misaligned set would settle the claim.","tokens_in":15577,"feed_emoji":"👥","tokens_out":6592,"duration_ms":62850,"temperature":0.7,"pith_summary":"This paper tries to establish that RGB-thermal image fusion, organized by two attention-style fusion modules and trained with a spatial-offset augmentation, makes UAV-view crowd counting and head localization more accurate than single-modality counting, especially at night or in dense scenes. The proposed network, TAPNet, reports mean absolute error 7.32 and F1 0.657 on DroneRGBT and 7.87 and 0.526 on GAIIC2 when both modalities are used, compared with larger errors for RGB-only or TIR-only versions. The claim matters because visible-light-only counters degrade under low light, and RGB-TIR cameras on drones are systematically misaligned; the paper's wager is that both problems can be handled at the feature-fusion level rather than by explicit image registration.","feed_headline":"Infrared plus visible cuts UAV crowd-count error","feed_subtitle":"A two-module attention fusion network reports MAE 7.32 and F1 0.657 on DroneRGBT, beating single-modality counting.","key_machinery":"The carrying mechanism is a pair of fusion modules plus a training-time alignment simulation. DAFP fuses the third and fourth backbon e feature maps of RGB and TIR by concatenating channel-pooled and spatially pooled features and generating per-channel and per-pixel weights through shared MLPs and convolutional layers. AFDF instead fuses early, running both modalities through a Transformer shared layer, a global-structure encoder, a detail encoder, and a hybrid Gaussian-Laplacian multi-kernel MMD term that pulls the modality distributions together in a reproducing kernel Hilbert space. Spatial random offset augmentation shifts the TIR image by random horizontal and vertical offsets in the range [-10, 10] during training, so the network sees misalignment similar to what the validation set contains.","core_discovery":"On the paper's own terms, the discovery is that a point-based head detection framework can be made multimodal by adding two complementary fusion paths: DAFP, which applies channel and spatial attention to the last two feature levels of RGB and TIR, and AFDF, an early-fusion encoder-decoder with a Transformer shared layer, global-structure and detail encoders, and hybrid multi-kernel MMD alignment. The paper reports that this combination lowers MAE on DroneRGBT from 10.32 for RGB-only and 13.15 for TIR-only to 7.32 for RGB-T input, and on GAIIC2 from 8.54 and 13.91 to 7.87, with corresponding F1 increases. It also reports that spatial random offset augmentation, shifting the TIR image by up to 10 pixels during training, reduces MAE by about 1.27 and 1.11 for the two fusion modules.","pith_inferences":["If the actual RGB-TIR misalignment in drone data is well modelled by a global translation of up to 10 pixels, then the augmentation is a near-free substitute for explicit registration; if misalignment is locally varying, AFDF's learned alignment would need to carry the full burden.","A natural testable extension is to apply the same hybrid-kernel MMD alignment to other paired modalities, such as RGB-depth or multi-view images, where systematic small shifts between sensors also occur.","Because GAIIC2's validation set was manually annotated by the authors, an independent re-annotation or a third-party benchmark would clarify how much of the reported gain is model behaviour versus annotation and matching protocol.","The reported F1 values are computed at threshold 0.8; evaluating across thresholds would show whether the localization gain is concentrated in high-confidence heads or uniform over the precision-recall curve."],"forward_implications":["With both modalities available, TAPNet reduces MAE on DroneRGBT from 10.32 to 7.32 and on GAIIC2 from 8.54 to 7.87, while raising F1, which makes RGB-TIR counting preferable whenever a thermal camera is available.","The auxiliary-point matching framework works best with only positive auxiliary points on fused bi-optical data, suggesting that fused features make negative-point guidance less necessary in this regime.","Head-point counting outperforms box counting by 1.66 MAE and 0.032 F1 on DroneRGBT, so point supervision is the more effective label type for dense UAV crowds.","Spatial random offset augmentation is a cheap route to misalignment robustness, improving MAE by 1.27 for DAFP and 1.11 for AFDF without adding fusion machinery.","Early fusion via AFDF and late attention fusion via DAFP trade off: DAFP is faster and lighter, while AFDF gives higher F1 on the DroneRGBT evaluation, so the choice depends on deployment constraints."],"supporting_citations":[{"why":"Supplies the auxiliary-point guidance matching framework that TAPNet extends to multimodal input.","marker":"[2]"},{"why":"Defines the point-based counting and localization baseline and the proposal-matching paradigm used for comparison.","marker":"[31]"},{"why":"Provides the DroneRGBT RGB-TIR drone dataset and benchmark used for the main evaluation.","marker":"[26]"},{"why":"Documents that RGB-TIR pairs from sensors with different fields of view and timestamps are systematically misaligned, motivating AFDF and the offset augmentation.","marker":"[42]"},{"why":"The long-range attention fusion network that inspires the dual-attention fusion module DAFP.","marker":"[6]"},{"why":"The dual-branch feature decomposition and domain-adaptive fusion work that inspires the AFDF module.","marker":"[37]"},{"why":"The Hungarian algorithm is the assignment method used for one-to-one proposal-target matching.","marker":"[13]"},{"why":"Supplies the large-scale jitter training strategy adopted for data augmentation.","marker":"[4]"},{"why":"Supplies the copy-paste augmentation context for the large-scale jitter and random scaling schedule.","marker":"[8]"}],"fun_headline_variants":["Fusing IR and RGB sharpens drone crowd counts","IR-RGB attention fusion boosts UAV crowd counting","TAPNet fuses infrared and visible to count crowds","Dual-optical attention network cuts dense crowd errors","Infrared adds edge for crowd head counting in dark"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The spatial random offset augmentation assumes that the systematic misalignment between RGB and TIR pairs is a global translation of at most 10 pixels, and that jittering the thermal image by that amount makes training images match the validation distribution; the paper does not measure the actual misalignment or test other shift magnitudes.","fun_headline_variants_meta":{"raw":{"variants":["Fusing IR and RGB sharpens drone crowd counts","IR-RGB attention fusion boosts UAV crowd counting","TAPNet fuses infrared and visible to count crowds","Dual-optical attention network cuts dense crowd errors","Infrared adds edge for crowd head counting in dark"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000184,"raw_usage":{"total_tokens":1306,"prompt_tokens":918,"completion_tokens":388,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":312}},"tokens_in":534,"tokens_out":388,"duration_ms":4379,"temperature":1.0,"reasoning_tokens":312,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:29:33.426211+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the actual per-pair displacement between RGB and TIR in DroneRGBT and GAIIC2, for example by matching annotated head points or estimating homographies: if typical misalignment exceeds 10 pixels or varies across the image, the augmentation's assumed distribution is wrong, and the reported gains should not transfer to other misaligned datasets. Retraining with offsets drawn from a wider or adaptive distribution and comparing MAE on an independent misaligned set would settle the claim.","supporting_citations":[{"cited_title":"Rethinking counting and localization in crowds: A purely point-based framework","cited_arxiv_id":null,"evidence_quote":"Defines the point-based counting and localization baseline and the proposal-matching paradigm used for comparison."},{"cited_title":"Rgb-t crowd counting from drone: A benchmark and mmccn network","cited_arxiv_id":null,"evidence_quote":"Provides the DroneRGBT RGB-TIR drone dataset and benchmark used for the main evaluation."},{"cited_title":"C²former: Calibrated and complementary transformer for rgb-infrared object de- tection","cited_arxiv_id":null,"evidence_quote":"Documents that RGB-TIR pairs from sensors with different fields of view and timestamps are systematically misaligned, motivating AFDF and the offset augmentation."},{"cited_title":"Lraf-net: Long-range attention fusion network for visible–infrared object detection.IEEE Transac- tions on Neural Networks and Learning Systems, 35:13232– 13243, 2023","cited_arxiv_id":null,"evidence_quote":"The long-range attention fusion network that inspires the dual-attention fusion module DAFP."},{"cited_title":"DAF-Net: A Dual-Branch Feature Decomposition Fusion Network with Domain Adaptive for Infrared and Visible Image Fusion","cited_arxiv_id":"2409.11642","evidence_quote":"The dual-branch feature decomposition and domain-adaptive fusion work that inspires the AFDF module."},{"cited_title":"The hungarian method for the assignment problem","cited_arxiv_id":null,"evidence_quote":"The Hungarian algorithm is the assignment method used for one-to-one proposal-target matching."},{"cited_title":"Simple Training Strategies and Model Scaling for Object Detection","cited_arxiv_id":"2107.00057","evidence_quote":"Supplies the large-scale jitter training strategy adopted for data augmentation."},{"cited_title":"Simple copy-paste is a strong data augmentation method for instance segmentation","cited_arxiv_id":null,"evidence_quote":"Supplies the copy-paste augmentation context for the large-scale jitter and random scaling schedule."}],"review_version":1}