{"id":"2bba9bf5-3409-4740-8476-d71174f609cd","arxiv_id":"2507.17995","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The first aerial-ground video person re-identification dataset with RGB and infrared modalities is introduced, along with a three-stream method that outperforms prior approaches on it.","lead":"This paper introduces AG-VPReID.VIR, a video person re-identification dataset that combines drone and ground cameras in both visible and infrared light, with 1,837 identities and about 125,000 frames. It also proposes a three-stream method that reports large gains on this dataset, but the paper's artifacts and evaluation choices need scrutiny.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported >20-point gains are not trustworthy: Appendix Sec. 10 selects loss hyperparameters λ1–λ4 by maximizing performance on a test protocol (Aerial→Ground I2V), so the final model is tuned to the test set and the comparison with baselines is biased.","rationale":"I read the paper as having two contributions: a new dataset and a new method. The dataset's label correctness is a genuine risk, but the most load-bearing insecurity for the stated performance claim is the evaluation protocol: the paper itself shows hyperparameters were tuned on a test protocol (Appendix Sec. 10, Table 9). This is not an external disagreement; it is an internal protocol violation. Even if all labels are perfect, the headline 'over 20 points' improvement cannot be taken at face value because the final model was selected using test labels. I therefore cannot accept the method claim as presented. The reader's CONDITIONAL verdict remains appropriate, and the release of code/data plus a validation-split rerun would settle the question. I agree with the reader that label verification is important, but I would weight the test-set hyperparameter selection as the single most concrete threat to the central claim. My agreement with the reader's weakest_assumption is partial for this reason.","tokens_in":15467,"tokens_out":9043,"duration_ms":98640,"concrete_test":"Ask the authors to rerun the full TCC-VPReID pipeline with λ1–λ4 selected on a held-out validation split disjoint from all four test protocols (for example, split the 326 training IDs into train/val) and report the same tables. If Rank-1 on Aerial→Ground I2V or Ground→Ground I2V drops by more than a few points absolute (e.g., >5) relative to the test-selected values, the reported gains are substantially inflated. Also report mean±std over at least three seeds with fixed hyperparameters.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The method claim in the abstract and Section 5.3—that TCC-VPReID outperforms prior methods by >20 points on AG-VPReID.VIR—rests on the evaluation being an unbiased estimate of generalization. That condition is violated. Appendix Sec. 10 (Table 9) reports a hyperparameter search over λ1–λ4 in L_total, conducted on the Aerial→Ground I2V protocol. Table 3 lists Aerial→Ground as one of the four test protocols. The values chosen by this search (λ1=1.0, λ2=1.5, λ3=1.0, λ4=1.5) are then fixed in Section 5.2 and used for all reported results, including the headline Ground→Ground numbers in Table 4. Thus the model has effectively been selected on a test split. Competing methods (SAADG, CST, AuxNet, etc.) were not given the same test-set tuning opportunity, so the reported 20.48–35.11 point Rank-1 gaps conflate algorithmic merit with test-set overfitting. No validation split, multiple-seed variance, or error bars are reported, so the magnitude of the bias cannot be assessed from the paper. This concern is directly evidenced by the manuscript itself, rather than being a speculation about data quality. The dataset contribution may still be valid, but the central method-superiority claim is not established by the numbers as reported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces AG-VPReID.VIR, claimed to be the first aerial-ground cross-modality video-based person Re-ID dataset, with 1,837 identities, 4,861 tracklets, and 124,855 frames collected from UAV, CCTV, and wearable cameras in RGB and IR. It also proposes TCC-VPReID, a three-stream architecture combining style-robust feature learning, memory-based cross-view adaptation, and intermediary-guided temporal learning, and reports benchmarks on HITSZ-VCM, BUPTCampus, and AG-VPReID.VIR. The paper claims large performance gains over prior methods, especially on AG-VPReID.VIR; however, the evaluation methodology contains a significant selection-bias problem because the loss hyperparameters are tuned on one of the test protocols.","tokens_in":15823,"tokens_out":4442,"duration_ms":49727,"significance":"If the dataset is released with reliable labels, it is a valuable community resource: it is the first dataset combining aerial and ground platforms with RGB-IR modalities for video Re-ID, and it addresses a real gap in the literature. The paper also provides a broad comparison with many existing methods, ablations of the three streams, qualitative results, and explicit ethics procedures. The method's reported superiority, however, is not established by the current experiments because the model is selected on a test protocol and results are reported from single runs, so the practical significance of the proposed architecture remains uncertain.","major_comments":[{"comment":"The loss hyperparameters λ1–λ4 are selected by maximizing performance on the Aerial→Ground I2V protocol (Table 9), which is explicitly a test protocol in Table 3. The chosen values (λ1=1.0, λ2=1.5, λ3=1.0, λ4=1.5) are then fixed in Sec. 5.2 and used for all reported results, including the headline Ground↔Ground numbers in Table 4. This means the model has been tuned to a test split, while the compared baselines were not given the same tuning opportunity. The reported 20+ point gains therefore conflate algorithmic merit with test-set overfitting and are not unbiased estimates of generalization. The authors must retune hyperparameters on a held-out validation set disjoint from all four test protocols, or explicitly justify why Aerial→Ground I2V is not part of the evaluation used to report final results.","section":"Sec. 10, Table 9; Sec. 5.2; Table 4"},{"comment":"All results are reported from single runs with no standard deviations, confidence intervals, or multiple random seeds. The central claim is a performance gap of more than 20 Rank-1 points over prior methods; without variance estimates, it is impossible to assess whether this gap is stable or an artifact of a particular initialization or data split. The authors should report mean and standard deviation over at least three random seeds for their method and ideally for the main baselines.","section":"Sec. 5.3, Table 4"},{"comment":"Table 8 lists architectural components that do not appear in the method equations, including Cross-View and Cross-Modal Graph Interaction, Domain-Adversarial Alignment, Temporal Memory Diffusion, Modality Decoupling, Multi-level Feature Integration, and Cross-stream Knowledge Transfer. None of these components is defined in Sec. 4, and the ablations in Tables 5 and 6 only vary St1, St2, and St3. It is therefore unclear which components are actually active in the model that produces the results in Table 4. The authors should either provide precise definitions and equations for all components claimed to be part of TCC-VPReID, or remove the unimplemented terms from Table 8 so that the described architecture matches the evaluated architecture.","section":"Appendix Table 8 vs. Sec. 4.1–4.4"},{"comment":"The dataset labels are the load-bearing foundation of every benchmark number, but the labeling description relies on 'manual review' and 'cross-verification' without any quantitative evidence of label quality. No inter-annotator agreement is reported, no protocol for resolving disagreements is described, and no independent verification subset is provided. If any tracklet mixes identities or if RGB and IR labels are misaligned, all reported results on AG-VPReID.VIR are affected. The authors should report annotation consistency statistics, release an independently labeled subset, or otherwise demonstrate label reliability across modalities and platforms.","section":"Sec. 3.2"}],"minor_comments":[{"comment":"The text states that V2I improvements over the best competitors are '33.69–35.11% in Rank-1', but Table 4 shows differences of 33.44, 33.69, and 34.51 against AuxNet, CST, and SAADG respectively (or 35.12 against SGIEL); the range should be corrected.","section":"Sec. 5.3"},{"comment":"The 'anaglyph representation' is described as the output of edge detection with an offset k, but k is never defined and the term 'anaglyph' usually refers to a color-channel composite, not an edge-filter output. Please clarify the operator A, the offset, and why this representation is called anaglyph.","section":"Eq. (9)"},{"comment":"There are frequent spacing inconsistencies such as 'UA V' instead of 'UAV', and Table 1 contains the typo 'HITZS-VCM' instead of 'HITSZ-VCM'. These should be cleaned up throughout.","section":"Sec. 3.1 and Table 1"},{"comment":"The platform-specific prompt P^m_{y_i} and the 'dual-branch attention-based decoder' are not defined in enough detail to be implemented; please specify the form of P, how it is generated, and how Eq. (7) is trained.","section":"Sec. 4.2, Eq. (7)"},{"comment":"The footnote states that 1,184 distractor identities are added to the gallery for all I2V experiments, but the rows for I2V do not show these distractors in the tracklet/frame counts. Please clarify whether the reported Q/G counts include or exclude the distractor set.","section":"Table 3"}],"recommendation":"major_revision","confidential_remarks":"The dataset contribution is genuinely promising, and I would not reject the paper on the basis of the current issues. The central problem is that the method claim is not yet supported by an unbiased evaluation: hyperparameters are selected on a test protocol, and no variance information is reported. These issues are fixable within the scope of the paper. I would also urge the editor to ensure that the architectural description in Appendix Table 8 is made consistent with the actual method, since the current discrepancy undermines reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The dataset is the real contribution here. As far as I can tell, AG-VPReID.VIR is the first video Re-ID benchmark to combine aerial and ground platforms with RGB and IR modalities, and that combination fills a genuine gap for 24-hour surveillance research. The collection effort looks serious: 1,837 identities, 4,861 tracklets, manual annotation with cross-verification, ethics approval, and a sensible train/test split. If the labels hold up, this is a useful resource for the community.\n\nThe method, TCC-VPReID, is honestly described as a recombination of known components: style augmentation, memory networks, anaglyph/edge-based intermediate representations. That is fine for a baseline, but the paper sells it as a novel architecture with massive gains, and that is where the problems start.\n\nThe stress-test concern is real and it lands. Appendix Section 10 explicitly tunes the loss weights λ1–λ4 on the Aerial→Ground I2V protocol, and Table 3 lists Aerial→Ground as one of the four test protocols. Those same weights are then used for all reported results, including the Ground→Ground numbers in Table 4. So the model has effectively been selected on a test split. The baselines were not given the same test-set tuning opportunity, which means the claimed 20–35 point Rank-1 gaps are not an unbiased estimate of algorithmic merit. This is selection bias, not a conspiracy, but it does undermine the central method-superiority claim.\n\nOther issues are more minor but worth noting: results are single runs with no error bars, so we cannot tell whether the 2–3 point gains on HITSZ-VCM and BUPTCampus are meaningful. The appendix architecture table (Table 8) lists components like graph interaction, domain-adversarial alignment, and temporal memory diffusion that do not appear in the method equations in the main text. And the dataset itself is not yet released, so we cannot independently verify the labels or the tracklet counts.\n\nFor a reader, the takeaway is: use this paper for the benchmark, not for the method numbers. The dataset deserves serious referee time, but the authors should be asked to move hyperparameter selection to a validation split, report variance across runs, clarify the architecture table, and release the data and code before publication. I would accept it for peer review with those conditions, but I would not cite the method results as evidence.\n\nWorth a reading group discussion, mostly as a case study on how test-set tuning creeps into benchmark papers.","headline":"The dataset is a genuinely useful first, but the method's headline gains are inflated by test-set hyperparameter selection, so treat the benchmark numbers with caution.","tokens_in":16353,"tokens_out":1367,"would_cite":true,"duration_ms":17038,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper introduces AG-VPReID.VIR, the first video-based person re-identification dataset combining aerial and ground platforms with visible and infrared modalities, along with a three-stream method that outperforms prior approaches by…","keywords":["person re-identification","visible-infrared re-identification","aerial-ground surveillance","UAV imagery","thermal imaging","video-based re-identification","cross-modality benchmark","drone and CCTV"],"falsifier":"Re-annotate a random sample of tracklets—say 200 identities spanning all five camera views—by independent annotators without access to the original labels, and measure per-frame and per-tracklet identity agreement; if agreement falls below a high threshold (e.g., 95%) where the original labels disagree, the benchmark's numbers cannot be trusted. A cheaper partial check is to train a simple baseline on a test split where labels are manually verified and see whether the reported 20-point gains persist.","tokens_in":15280,"feed_emoji":"🚁","tokens_out":7367,"duration_ms":69495,"temperature":0.7,"pith_summary":"Person re-identification systems need to match the same person across cameras, but existing RGB-infrared benchmarks only cover ground-level views, and existing drone-based benchmarks only cover visible light. This paper introduces AG-VPReID.VIR, a video dataset of 1,837 identities in 4,861 tracklets and 124,855 frames, where the same people appear in RGB and infrared images from UAVs, fixed CCTV cameras, and wearable cameras. The authors argue that this combination—cross-platform and cross-modality at once—creates new and harder failure modes, and they report that established methods drop by more than half their accuracy on it. They also propose TCC-VPReID, a three-stream architecture that combines style-robust features, memory-based cross-view adaptation, and an intermediary-guided temporal stream, and report Rank-1/mAP of 36.18/41.56 for infrared-to-visible and 46.33/59.23 for visible-to-infrared ground-to-ground retrieval, more than 20 points above prior methods.","feed_headline":"First aerial-ground RGB-IR video re-ID benchmark set","feed_subtitle":"1,837 people appear in drone, CCTV, and wearable views in visible and thermal light; a three-stream model jumps over 20 points ahead of…","key_machinery":"The load-bearing object is the AG-VPReID.VIR dataset: 4,861 manually-reviewed tracklets of 1,837 identities, each identity labeled consistently across RGB and IR images from UAV, CCTV, and wearable cameras, with synchronized dual-sensor drone captures producing the first aerial infrared imagery for person re-identification. The method that carries the experimental claim is TCC-VPReID, a three-stream architecture: Stream 1 applies random channel-style perturbations plus intra-modal style attacks to force style-robust features; Stream 2 maintains separate aerial and ground memory prototypes per identity and aligns video features to them with a contrastive loss; Stream 3 converts RGB and IR frames into edge-based anaglyph images that serve as a modality-neutral intermediary, with cross-reconstruction pulling the two modalities together. The three streams are fused and trained with a weighted sum of identity, triplet, style-attack, cross-reconstruction, and video-to-memory losses.","core_discovery":"The paper's central claim is that person re-identification can and should be studied jointly across platforms (aerial and ground) and modalities (visible and infrared), and that AG-VPReID.VIR is the first dataset to make that possible. The dataset's unique value is aerial infrared imagery, which is harder than ground infrared because of small, low-resolution, top-down views of people. To show the setting is learnable, the paper builds TCC-VPReID, whose three streams each target one gap: the first attacks style variation with augmented channel perturbations, the second stores separate aerial and ground prototypes per identity in a memory bank with a video-to-memory contrastive loss, and the third uses edge-detected anaglyph images as a modality-neutral intermediary with cross-reconstruction. On the new benchmark the method lifts Rank-1 by more than 20 points over the best prior work, and on two existing ground-only video benchmarks it also improves on the previous state of the art.","pith_inferences":["Aerial infrared may be the main bottleneck: at 15–45 meters altitude with 640×512 resolution, a person's thermal signature is only a few dozen pixels, so the reported 20-point gains may partly reflect how quickly a model learns to exploit clothing-independent thermal silhouettes—an exploitable shortcut rather than identity understanding.","Because identities were collected over five months with consent and facial anonymization, the dataset could enable longitudinal studies of appearance change across days and seasons, which current single-day benchmarks do not capture.","The memory-based cross-view stream suggests a testable extension: using the same dual-memory idea to align other domain splits, such as day versus night or occluded versus unobstructed, without retraining.","A practical consequence the authors do not state: a drone-and-thermal-CCTV setup of this kind could allow nighttime search-and-rescue or security operations to match a ground witness description to an aerial thermal sighting, but only if precision at high ranks is much better than the reported roughly 36% Rank-1."],"forward_implications":["If the dataset is sound, it becomes the standard benchmark for aerial-ground RGB-IR video re-identification, letting researchers compare methods under a realistic 24-hour surveillance mix of drones, CCTV, and body cameras.","The reported sharp performance drop of existing methods when moving to this dataset indicates that cross-platform variation, not just modality gap, is a major and understudied source of difficulty.","The V2I/I2V asymmetry (visible queries fare much better than infrared queries) implies that thermal imagery loses identity information that visible images retain, motivating work on thermal image enhancement or query-side fusion.","TCC-VPReID's three-stream design gives a reusable template: style augmentation, per-view memory, and an intermediary modality can each be swapped or ablated in future work."],"supporting_citations":[{"why":"Provides the aerial-ground RGB image collection (UAV, CCTV, wearable) that inspired the platform mix in AG-VPReID.VIR.","marker":"[25]"},{"why":"Is the prior aerial-ground video dataset whose RGB-only scope the new dataset extends to infrared.","marker":"[47]"},{"why":"Contributes the HITSZ-VCM benchmark on which the method is evaluated and a memory-based temporal baseline.","marker":"[18]"},{"why":"Contributes the BUPTCampus video RGB-IR benchmark and the AuxNet baseline compared in experiments.","marker":"[6]"},{"why":"Provides the paired RGB-thermal image benchmark RegDB used as an early cross-modality reference.","marker":"[23]"},{"why":"Provides the SYSU-MM01 benchmark, the canonical RGB-IR re-identification dataset whose ground-only setting the new dataset contrasts with.","marker":"[34]"},{"why":"Supplies the style augmentation technique and the SAADG baseline that the new method outperforms by over 20 points.","marker":"[51]"},{"why":"Supplies the edge-detection anaglyph idea used in Stream 3 and the IBAN baseline compared in Table 4.","marker":"[14]"},{"why":"Supplies the pretrained object detector used to generate the initial person tracks that annotators reviewed.","marker":"[13]"}],"fun_headline_variants":["First UAV+CCTV RGB-IR video Re-ID dataset, 1,837 IDs","Aerial-ground thermal Re-ID: new benchmark, +20 Rank-1","Drones and fixed cams: new video Re-ID across RGB and IR","First cross-platform RGB-IR Re-ID video benchmark unveiled","New Re-ID dataset bridges aerial and ground, visible and thermal"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The dataset labels—the identity of every person in every RGB and IR tracklet, across aerial, CCTV, and wearable views—must be correct and consistent; if any tracklet mixes identities or any visible and infrared label is misaligned, every accuracy number on AG-VPReID.VIR is compromised.","fun_headline_variants_meta":{"raw":{"variants":["First UAV+CCTV RGB-IR video Re-ID dataset, 1,837 IDs","Aerial-ground thermal Re-ID: new benchmark, +20 Rank-1","Drones and fixed cams: new video Re-ID across RGB and IR","First cross-platform RGB-IR Re-ID video benchmark unveiled","New Re-ID dataset bridges aerial and ground, visible and thermal"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1334,"prompt_tokens":1011,"completion_tokens":323,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":229}},"tokens_in":627,"tokens_out":323,"duration_ms":3968,"temperature":1.0,"reasoning_tokens":229,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:38:38.876151+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-annotate a random sample of tracklets—say 200 identities spanning all five camera views—by independent annotators without access to the original labels, and measure per-frame and per-tracklet identity agreement; if agreement falls below a high threshold (e.g., 95%) where the original labels disagree, the benchmark's numbers cannot be trusted. A cheaper partial check is to train a simple baseline on a test split where labels are manually verified and see whether the reported 20-point gains persist.","supporting_citations":[{"cited_title":"Nguyen, K","cited_arxiv_id":null,"evidence_quote":"Provides the aerial-ground RGB image collection (UAV, CCTV, wearable) that inspired the platform mix in AG-VPReID.VIR."},{"cited_title":"Zhang, W","cited_arxiv_id":null,"evidence_quote":"Is the prior aerial-ground video dataset whose RGB-only scope the new dataset extends to infrared."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the HITSZ-VCM benchmark on which the method is evaluated and a memory-based temporal baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the BUPTCampus video RGB-IR benchmark and the AuxNet baseline compared in experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the paired RGB-thermal image benchmark RegDB used as an early cross-modality reference."},{"cited_title":"Wu, W.-S","cited_arxiv_id":null,"evidence_quote":"Provides the SYSU-MM01 benchmark, the canonical RGB-IR re-identification dataset whose ground-only setting the new dataset contrasts with."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the style augmentation technique and the SAADG baseline that the new method outperforms by over 20 points."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the edge-detection anaglyph idea used in Stream 3 and the IBAN baseline compared in Table 4."},{"cited_title":"Jocher, Ayush, and J","cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained object detector used to generate the initial person tracks that annotators reviewed."}],"review_version":1}