{"id":"895286e6-c81c-4802-91f1-90e39483e0c2","arxiv_id":"1908.10049","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"GLTR combines dilated temporal pyramid convolution and temporal self-attention over frame features to improve video person re-identification accuracy across five datasets.","lead":"This paper proposes a video representation for person re-identification that combines short-range temporal cues from dilated temporal convolutions with long-range relations from a temporal self-attention module. The method is tested on five video person re-identification datasets, including a newly collected large-scale dataset, and reports competitive or state-of-the-art rank-1 accuracy.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The unqualified 'state-of-the-art' claim selects rank-1 as the decisive metric, while GLTR trails STA in mAP on both MARS and Duke; the rank-1 advantage is small and unreplicated.","rationale":"The reader identified metric selection as the weakest assumption, and the paper's own numbers support that concern: GLTR's mAP is lower than STA's on both MARS and Duke, while the rank-1 advantages are modest and unaccompanied by variance estimates. This does not undermine the core methodological claim that combining short-term dilated temporal convolution with long-term temporal self-attention improves video re-identification over the baseline and over several prior methods; the ablations in Table 2 and comparisons in Tables 4-7 show consistent gains from the proposed components. However, the headline state-of-the-art assertion is overstated because it depends on treating rank-1 as the single decisive metric, and the paper does not justify that choice. The reader's CONDITIONAL verdict is therefore appropriate: the contribution is credible, but the superiority claim should be qualified, and the evaluation would benefit from repeated runs and variance reporting. No additional load-bearing concern was found that would require rejection or a stronger downgrade.","tokens_in":13922,"tokens_out":4889,"duration_ms":55486,"concrete_test":"Retrain GLTR on MARS and DukeMTMC-VideoReID with at least five random seeds and report mean and standard deviation for both rank-1 and mAP under the same evaluation protocol. Then compare against the published STA results in Tables 5 and 7. If the rank-1 gap between GLTR and STA falls within one standard deviation, or if GLTR's mAP remains below STA's, the abstract's 'better than current state-of-the-art' claim should be qualified to 'competitive rank-1 accuracy with lower mAP than STA on some benchmarks'.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim 'better than current state-of-the-art' is supported only by rank-1 accuracy, and the paper's own comparison tables show that the claim is metric-dependent. On MARS (Table 5), GLTR achieves rank-1 87.02% versus STA's 86.30%, but mAP is 78.47% versus STA's 80.80%. On DukeMTMC-VideoReID (Table 7), GLTR rank-1 is 96.29% versus STA's 96.20%, but mAP is 93.74% versus STA's 94.90%. Under the mAP criterion, which the paper itself reports and uses for LS-VID, STA is better on both datasets. The rank-1 advantages are also small (0.72 and 0.09 percentage points), and the paper reports no repeated runs, error bars, or significance testing for these comparisons, so the measured rank-1 edge may be within run-to-run variation. The methodological contribution itself is supported by the ablations, where DTP and TSA each improve over baseline and their combination is consistently best; the load-bearing weakness is specifically the unqualified superiority statement in the abstract and introduction, not the proposed architecture.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Global-Local Temporal Representation (GLTR) for video person re-identification, formed by a Dilated Temporal Pyramid (DTP) convolution to capture short-term local temporal cues and a Temporal Self-Attention (TSA) model to capture long-term global relations across inconsecutive frames. The two modules are applied to frame-level features extracted from a ResNet50 backbone and aggregated by average pooling. The method is evaluated on four standard video ReID benchmarks (PRID, iLIDS-VID, MARS, DukeMTMC-VideoReID) and on a newly introduced large-scale dataset (LS-VID), with systematic ablations showing that each component improves over the baseline and their combination is generally best. The paper claims state-of-the-art rank-1 accuracy on MARS (87.02%) without re-ranking and reports competitive results on the other datasets.","tokens_in":14127,"tokens_out":6784,"duration_ms":67402,"significance":"The paper's contribution is an architecture that jointly models local and global temporal cues in a simple, single-stream CNN, and it provides a thorough empirical study across five datasets, including a new large-scale dataset (LS-VID). The ablation study is a strength: it isolates the contribution of DTP and TSA, compares DTP with alternative local temporal modeling strategies, and reports consistent gains over the average-pooling baseline. If the claims hold, the method is a practical, efficient alternative to more complex RNN or 3D-CNN designs. However, the headline claim of 'better than current state-of-the-art' is not fully substantiated because the comparison with STA shows that STA achieves higher mAP on MARS and DukeMTMC-VideoReID, and the reported rank-1 margins are small and without variance analysis.","major_comments":[{"comment":"The definition of dilated temporal convolution is inconsistent with the text. Eq. (5) computes f^{(r)}_t as the sum over i=1..w of f[t+r·i] × W[i], which yields a one-sided (causal) sampling of future frames. However, the text states that with r=2, w=3, 'each output feature corresponds to a temporal range of five adjacent frames,' which implies a centered kernel with offsets t−r, t, t+r. The formula as written would cover frames t+2, t+4, t+6, not a symmetric five-frame neighborhood. This inconsistency affects the reproducibility of the core DTP module and must be corrected, either by revising the equation to use centered offsets or by aligning the text with the actual implementation.","section":"Section 3.2, Eq. (5)"},{"comment":"The unqualified claim 'better than current state-of-the-art' in the abstract and introduction is not supported by the paper's own tables. On MARS (Table 5), GLTR achieves rank-1 87.02% vs STA's 86.30%, but mAP is 78.47% vs STA's 80.80%. On DukeMTMC-VideoReID (Table 7), GLTR achieves rank-1 96.29% vs STA's 96.20%, but mAP is 93.74% vs STA's 94.90%. Thus, if mAP is taken as an equal or primary metric, GLTR is not superior to STA on either dataset. The claim should be qualified explicitly (e.g., 'state-of-the-art rank-1 accuracy') or the comparison should be presented in a metric-neutral way.","section":"Abstract, Introduction, Tables 5 and 7"},{"comment":"The rank-1 improvements over STA are small (0.72 percentage points on MARS and 0.09 percentage points on DukeMTMC-VideoReID), and the paper reports no repeated trials, error bars, or significance tests for these datasets. For PRID and iLIDS-VID, the authors average over 10 random splits, showing awareness of run-to-run variance, but no such variance is reported for MARS or Duke. The measured rank-1 edges may be within normal variation, and the paper should either provide variance estimates or temper the claim of superiority from these small margins.","section":"Tables 5 and 7"}],"minor_comments":[{"comment":"The phrase 'state-of-the art' is missing a hyphen; it should be 'state-of-the-art'.","section":"Abstract"},{"comment":"The text contains a duplicated word: 'generalize self-attention to to capture' should be 'generalize self-attention to capture'.","section":"Section 3.3"},{"comment":"The caption contains typographical errors: 'first tow' and 'second tow' should be 'first two' and 'second two'.","section":"Figure 7 caption"},{"comment":"The sentence 'We repeat PBR [39] and DRSA [22] by referring to their implantations on MARS' uses 'implantations'; it should be 'implementations'.","section":"Section 4.4"},{"comment":"The text describes a T×T attention mask M but does not mention the softmax normalization that is shown in Figure 2. Please specify the normalization step explicitly in the equation or following text.","section":"Section 3.3, Eq. (7)"},{"comment":"The paper does not state whether LS-VID will be released publicly. Given that the dataset is a claimed contribution, a statement about its availability (or a URL) is needed for reproducibility.","section":"Section 4.1, LS-VID"}],"recommendation":"major_revision","confidential_remarks":"The architectural contribution is solid, and the ablation study provides good evidence for the effectiveness of DTP and TSA over the baseline. The main obstacles to acceptance are the overstated state-of-the-art claim and the technical error in the DTP formulation. Both are fixable in revision; the paper should also address the small margin and lack of error bars in the STA comparison. If the authors revise the abstract and introduction to be metric-selective, correct Eq. (5), and add variance or qualification, the paper could be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has one real contribution and one real flaw. The contribution is a clean, cheap temporal aggregation module for video person ReID: dilated temporal pyramid convolutions for local motion cues plus temporal self-attention to down-weight occluded frames, all on top of a single-stream ResNet50 with softmax loss. The ablations on five datasets are consistent: each component beats the baseline, and the combination beats both. That evidence is credible and it is the heart of the paper. The flaw is the headline claim. The abstract says 'better than current state-of-the-art' based on rank-1 on MARS (87.02 vs 86.30 for STA), but on the same table STA has higher mAP (80.80 vs 78.47). On Duke the pattern repeats: rank-1 edges 96.29 to 96.20 but mAP loses 93.74 to 94.90. The rank-1 margins are 0.72 and 0.09 points, with no error bars or significance testing, so the superiority claim is both metric-selected and numerically fragile. That needs to be fixed in any revision.\n\nThe design itself is sound. Dilated convolution is borrowed from segmentation and self-attention from action recognition and image ReID, so the novelty is in the combination and the systematic comparison rather than the ingredients. The paper also introduces LS-VID, a 3,772-identity, 15-camera dataset, which could be a useful benchmark, but it is not released and neither is the code, so external verification is limited. The LS-VID controlled comparisons against re-implementations of previous methods are the strongest part of the evaluation; the standard-benchmark comparisons rely on published numbers, which is normal but weaker.\n\nThe free parameters (number of pyramid branches, attention compression factor) are chosen on validation sets, which is standard model selection, not circularity. The paper is honestly written, the methodology section is clear, and the empirical support for the architecture is genuine. My main worry is that the overstated SOTA claim is what a casual reader will remember.\n\nThis paper deserves a serious referee. The experiments are reproducible in principle, the ablations are informative, and the method is simple enough to be useful. I would accept it with a requested revision to qualify the SOTA language and to report variance or at least multiple runs on the key comparisons. It is a reasonable paper, not a breakthrough.","headline":"Sound empirical work with a clean two-part temporal aggregation module; the only load-bearing flaw is a rank-1-only SOTA claim that doesn't survive mAP on two benchmarks.","tokens_in":14674,"tokens_out":3468,"would_cite":false,"duration_ms":32559,"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":"The paper claims that a video-based person re-identification system can reach state-of-the-art rank-1 accuracy by combining a dilated temporal pyramid convolution for local motion cues with a temporal self-attention module for global…","keywords":["video person re-identification","temporal representation","dilated temporal convolution","temporal self-attention","multi-scale temporal cues","occlusion robustness","LS-VID dataset","video sequence aggregation"],"falsifier":"Run the standard MARS and DukeMTMC-VideoReID evaluations with mean average precision as the primary metric. On both datasets the paper's own tables show the competing STA method reporting higher mAP than GLTR (80.80 vs 78.47 on MARS; 94.90 vs 93.74 on Duke), so under a mAP-primary protocol the claim that GLTR beats current state of the art fails. A reader can settle the question by reproducing both models under identical backbone, loss, and evaluation settings.","tokens_in":13689,"feed_emoji":"🎥","tokens_out":8638,"duration_ms":82502,"temperature":0.7,"pith_summary":"This paper is trying to establish that a single-stream video representation, built by stacking two temporal modules on top of ordinary frame features, captures what video person re-identification needs: local cues from adjacent frames (motion, gait, speed) and global cues from distant frames (occlusion and noise context). The proposed Global-Local Temporal Representation (GLTR) applies Dilated Temporal Pyramid (DTP) convolution followed by Temporal Self-Attention (TSA), then average-pools the result. The authors argue this simple design outperforms temporal pooling, recurrent networks, 3D convolutions, and body-part-cue systems, reporting rank-1 accuracy of 87.02% on MARS without re-ranking and strong results on PRID, iLIDS-VID, DukeMTMC-VideoReID, and their new LS-VID dataset. A sympathetic reading is that the combination of local and global temporal modeling is the active ingredient, since each module alone improves the baseline and together they improve it further.","feed_headline":"Temporal cues lift video person Re-ID to 87.02% on MARS","feed_subtitle":"Short-range motion and long-range attention combine in one stream, beating complex body-part and 3D models","key_machinery":"The load-bearing object is the aggregation subnetwork that turns a sequence of $d$-dimensional frame features into the GLTR vector. It contains two named modules. DTP runs $N=3$ parallel dilated temporal convolutions with dilation rates $1,2,4$ along the frame index, so each output frame mixes a growing window of neighbors without increasing parameters or downsampling; this encodes short-term motion and appearance cues. TSA projects the DTP output through two convolution layers, multiplies one projection by the transpose of the other to form a $T\\times T$ attention mask, and re-weights the feature sequence with a residual connection whose final convolution is zero-initialized; this lets distant, non-adjacent frames vote on each other and suppresses occluded frames. Average pooling over the updated sequence produces the fixed-length GLTR. The machinery's work is to fuse two complementary temporal scales so the final representation is both discriminative, through local motion, and robust, through global context.","core_discovery":"The paper's central claim is that the right video representation for person re-identification must jointly encode local temporal cues, such as the motion pattern and speed visible across adjacent frames, and global temporal cues, such as the relations among distant frames that reveal occlusions and noises. It constructs GLTR by applying Dilated Temporal Pyramid (DTP) convolution followed by Temporal Self-Attention (TSA) to a sequence of frame features, then average-pooling the updated features. The paper reports that this representation reaches rank-1 accuracy of 87.02% on MARS without re-ranking, outperforming recent systems that use body-part cues, recurrent networks, or 3D convolutions, and that it is consistently competitive on PRID, iLIDS-VID, DukeMTMC-VideoReID, and the newly collected LS-VID dataset.","pith_inferences":["Editorial inference: the state-of-the-art claim is metric-dependent, since the paper's own tables show the competing STA method with higher mAP on MARS (80.80 vs 78.47) and on DukeMTMC-VideoReID (94.90 vs 93.74).","Editorial inference: the same two-module recipe could transfer to other video understanding tasks where short-range motion and long-range context matter, because the aggregation subnetwork does not depend on re-identification-specific losses or part models.","Editorial inference: the learned attention weights could be reused as a frame-quality or occlusion score to filter training samples or to explain why a match succeeded, which the paper visualizes in its occlusion example.","Editorial inference: a natural test is to vary inference-time sequence length; because the $T\\times T$ attention mask has constant cost per pair of frames, the method should be especially favorable when sequences grow."],"forward_implications":["A video re-identification system can get competitive or state-of-the-art rank-1 accuracy without body-part supervision, optical flow, or a recurrent or 3D-convolutional head, using only frame features plus the two modules.","Because the aggregation subnetwork is backbone-agnostic, stronger image-based re-identification features can be plugged in directly and should further improve GLTR.","The attention mask provides an explicit mechanism for occlusion robustness, so tracklets with noisy or partially occluded frames are down-weighted during pooling rather than corrupting the video descriptor.","On longer and more varied sequences such as LS-VID, the gain over recurrent and 3D models is several rank-1 points, suggesting the local-plus-global design scales with sequence length and dataset difficulty."],"supporting_citations":[{"why":"It supplies the dilated spatial pyramid idea that DTP adapts to the temporal axis.","marker":"[4, 51]"},{"why":"It supplies the dilated convolution formalism used to enlarge temporal receptive fields.","marker":"[52]"},{"why":"It provides the self-attention context-modeling basis that TSA generalizes to the temporal axis.","marker":"[10, 15, 53]"},{"why":"It provides the MARS benchmark and the baseline evaluation protocol.","marker":"[56]"},{"why":"It provides the DukeMTMC-VideoReID dataset and the ETAP-Net comparison.","marker":"[47, 34]"},{"why":"It is the strongest competing method on MARS and Duke, and its mAP numbers delimit the state-of-the-art claim.","marker":"[11]"},{"why":"It is the 3D CNN baseline that GLTR outperforms.","marker":"[19]"},{"why":"It is the recurrent baseline that GLTR outperforms.","marker":"[29]"},{"why":"It is the body-part-cue baseline that GLTR outperforms on MARS.","marker":"[39]"}],"fun_headline_variants":["GLTR: short and long temporal cues for 87.02% on MARS","One-stream model beats body-part and 3D net on Re-ID","Temporal pyramid plus attention hits 87% on MARS","Local motion and global attention: new Re-ID best","Dual time scales lift video person Re-ID to 87.02%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole \"better than state-of-the-art\" case rests on the choice of rank-1 accuracy as the headline metric; if a reader weighs mean average precision equally, the competitor STA still leads on MARS and DukeMTMC-VideoReID.","fun_headline_variants_meta":{"raw":{"variants":["GLTR: short and long temporal cues for 87.02% on MARS","One-stream model beats body-part and 3D net on Re-ID","Temporal pyramid plus attention hits 87% on MARS","Local motion and global attention: new Re-ID best","Dual time scales lift video person Re-ID to 87.02%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00021,"raw_usage":{"total_tokens":1375,"prompt_tokens":877,"completion_tokens":498,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":493,"completion_tokens_details":{"reasoning_tokens":404}},"tokens_in":493,"tokens_out":498,"duration_ms":5414,"temperature":1.0,"reasoning_tokens":404,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:53:36.379477+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the standard MARS and DukeMTMC-VideoReID evaluations with mean average precision as the primary metric. On both datasets the paper's own tables show the competing STA method reporting higher mAP than GLTR (80.80 vs 78.47 on MARS; 94.90 vs 93.74 on Duke), so under a mAP-primary protocol the claim that GLTR beats current state of the art fails. A reader can settle the question by reproducing both models under identical backbone, loss, and evaluation settings.","supporting_citations":[{"cited_title":"Sta: Spatial-temporal attention for large-scale video- based person re-identification","cited_arxiv_id":null,"evidence_quote":"It is the strongest competing method on MARS and Duke, and its mAP numbers delimit the state-of-the-art claim."},{"cited_title":"Multi- scale 3d convolution network for video based person re- identification","cited_arxiv_id":null,"evidence_quote":"It is the 3D CNN baseline that GLTR outperforms."},{"cited_title":"Spatial and temporal mutual promotion for video-based person re-identification","cited_arxiv_id":null,"evidence_quote":"It is the recurrent baseline that GLTR outperforms."},{"cited_title":"Part-aligned bilinear representations for per- son re-identification","cited_arxiv_id":null,"evidence_quote":"It is the body-part-cue baseline that GLTR outperforms on MARS."}],"review_version":1}