{"id":"f72f8820-3193-46d7-a1ff-8279d1ba2425","arxiv_id":"2412.16506","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"DSCA improves unsupervised domain adaptive person search by adding a perception-driven adaptive filter and cluster-proxy memory, reaching 80.2% mAP on CUHK-SYSU and 39.9% mAP on PRW.","lead":"A person-search model trained on labeled data in one camera domain is adapted to a new unlabeled domain by filtering noisy pseudo-bounding-boxes and cleaning cluster assignments at the instance level. The proposed DSCA framework reports the best results among unsupervised domain adaptive person-search methods on CUHK-SYSU and PRW and posts code.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PDAF's channel-magnitude threshold may suppress foreground features in target domains; the paper provides no per-feature evidence that it separates noise from identity information.","rationale":"I read the paper's central claim as empirical: DSCA with PDAF and CPR outperforms DAPS and DDAM on CUHK-SYSU and PRW. The least secure condition for that claim is the mechanism of PDAF, which the reader also identified. The channel-statistics threshold is derived from the same feature map it filters, so it is not a measure of foreground targetness; in background-dominated target frames, the threshold can be inflated and attenuate genuine person features. The paper's ablations show end-to-end drops when components are removed, but they cannot distinguish between removing noise and destroying signal. That distinction matters because the stated contribution is specifically noise filtering. The proposed test uses available ground-truth boxes to directly measure whether the filter's suppression rates support the foreground/background separation story. This does not change the reader's conditional verdict: the empirical claims are plausible and worth checking, but the mechanism evidence is currently missing.","tokens_in":12140,"tokens_out":8017,"duration_ms":73059,"concrete_test":"Instrument the released code with hooks after the SCF in each of the four backbone layers. On the CUHK-SYSU and PRW test sets (which have ground-truth boxes), compute per layer: (i) the fraction of spatial locations inside ground-truth person boxes whose activations satisfy |F| < τ and are therefore zeroed; (ii) the same fraction for background locations; (iii) the average margin |F| − τ at foreground locations. Repeat for both source and target domains. If the foreground zeroing rate is not materially lower than the background rate, or is higher in the target domain than in the source, then PDAF is attenuating valid identity features and the noise-filtering interpretation of the ablation gains is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is that the PDAF threshold τ = α ⊙ AvgPool(F) (Eq. 2), computed from pooled feature statistics, separates foreground person features from background/noise in each of the first four ResNet-50 layers and in both domains. This is not established. Since τ is a scaled channel mean, a target frame with large background regions will have an inflated per-channel threshold; the SCF in Eq. 6 will then set any activation with |x| < τ to zero, including low-magnitude but identity-bearing person features. The paper's ablations (Tables 2–3) and LayerCAM visualizations (Fig. 7) only measure end-to-end accuracy and coarse attention; they do not measure how many true-positive foreground locations are zeroed by the filter, nor whether the filter's suppression rate is higher in the target domain than in the source domain. Without this, the core causal claim—that PDAF removes noisy pseudo-boxes rather than merely acting as a regularizer or distorting features—remains unsupported, and the reported gains could be due to an unrelated mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Dual Self-Calibration (DSCA), an unsupervised domain-adaptive person search framework built on a DAPS-style end-to-end detector. It introduces (i) a Perception-Driven Adaptive Filter (PDAF), comprising a Perception-Driven Threshold (PDT) in Eq. (2) and a Self-Calibrating Filter (SCF) in Eq. (6), which suppresses backbone features below a channel-wise adaptive threshold; and (ii) a Cluster Proxy Representation (CPR), which replaces instance-level memory with cluster-level proxies using the updates in Eqs. (8)-(10). Experiments on CUHK-SYSU and PRW report 80.2% mAP / 81.7% top-1 and 39.9% mAP / 81.6% top-1, outperforming DAPS and DDAM and also exceeding several weakly supervised baselines. Ablations in Tables 2-3 and visualizations in Fig. 7 are used to argue that each component contributes to the final result.","tokens_in":12358,"tokens_out":6504,"duration_ms":58477,"significance":"The significance is moderate-to-high if the reported results hold: the method advances the UDA person search state of the art, especially on PRW, while simplifying training through one clustering per epoch. The code release and the end-to-end ablations for both modules on both benchmarks are strengths, and the paper makes no hidden circular assumptions: the modules and ablations are defined independently of the reported numbers. However, the empirical case is limited by single-seed runs, per-dataset epoch schedules, and the absence of direct evidence for the proposed filtering mechanism, so the central claims should be treated as promising rather than fully established.","major_comments":[{"comment":"The central causal claim that PDAF removes noisy pseudo-boxes rather than distorting true foreground features is not directly supported. The threshold τ = α ⊙ AvgPool(F) is a scaled channel mean; in a target frame with a large background area this threshold can exceed low-magnitude but identity-bearing activations, and the SCF in Eq. (6) zeros exactly those activations. The ablations in Tables 2-3 and the LayerCAM visualizations in Fig. 7 measure only end-to-end accuracy or coarse attention; they do not measure how many true-positive foreground locations are zeroed by the filter, nor whether the suppression rate differs between source and target domains. Please add per-layer statistics (e.g., the fraction of foreground versus background activations zeroed, or recall of pseudo-boxes before and after filtering) and, ideally, a control that applies the same nonlinearity with shuffled or fixed thresholds to rule out that PDAF acts merely as an unstructured regularizer.","section":"§3.2, Eqs. (2) and (6)"},{"comment":"Table 2 is not self-contained: rows use blank cells to indicate which component is removed, and the text refers to rows as #1-#4 while the table has no row numbers. As printed, the reader cannot reliably map each row to the corresponding ablated component, which undermines the ablation evidence for PDT, SCF, and CPR. Please replace the blank cells with explicit ✓/✗ markers and add row labels or row numbers, and correct the comma decimal in the 34.7 entry.","section":"§4.2, Table 2"},{"comment":"The headline improvement over the runner-up is 0.7 mAP on CUHK-SYSU and 3.2 mAP on PRW, but all results come from a single run with no error bars, and the epoch schedules are chosen per target dataset (7 pre-training + 13 joint epochs for PRW, 2 + 7 for CUHK-SYSU). This makes it difficult to distinguish genuine improvement from run-to-run variance or benchmark-specific tuning. Please report multiple seeds (at least three) with means and standard deviations, and either fix a shared schedule or provide a sensitivity analysis over epochs and the main hyperparameters (γ, m, and the contrastive temperature).","section":"§4.1 and §4.3, Table 1"},{"comment":"The online cluster proxy update is written as ck ← −(1−γ)ck + γf. If taken literally, the negative sign makes the stored proxy alternate in sign with each update and is inconsistent with the momentum-update description in the text and with the positive combination in Eq. (10). Please confirm the intended formula and correct the sign in Eq. (9), since this equation is load-bearing for the CPR module.","section":"§3.3, Eq. (9)"}],"minor_comments":[{"comment":"There are several naming inconsistencies: \"DSAC\" appears in the Method section, \"DCSA\" appears in the Experiments section, and the Introduction refers to \"The SCT\" where the Self-Calibrating Filter (SCF) is meant. Please unify the acronyms.","section":"Throughout"},{"comment":"The phrase \"panning operation\" for the soft threshold should be \"shifting\" or \"shrinking operation\" to describe the translation of the input by τ.","section":"§3.2, Eq. (4)"},{"comment":"The order n of the High-Order Soft Threshold is never stated in the experimental setup; Fig. 3 only says \"take n = 2 as an example.\" Please state the value of n used in all experiments and include an ablation over n, since the text motivates SCT as a remedy for gradient issues at larger n but never reports how n was chosen.","section":"§3.2, Eqs. (5)-(6)"},{"comment":"The qualitative evidence would be easier to judge if Fig. 6 had labeled axes with units, and if Fig. 7 showed the input image and attention map for the same scale in each row; the current composites are hard to compare.","section":"§4.2, Fig. 7 and Fig. 6"},{"comment":"The phrase \"comparable to or even exceeding the performance of some fully supervised methods\" is supported only against older supervised baselines such as OIM; the modern supervised methods (SeqNeXt, SEAS) remain clearly ahead. Please rephrase the claim to avoid implying competitiveness with current fully supervised person search.","section":"§4.3, Table 1"}],"recommendation":"major_revision","confidential_remarks":"The main risk is empirical: on CUHK-SYSU the gain over DDAM is only 0.7 mAP, all runs are single-seed, and the PDAF mechanism is not directly validated. I would want multi-seed results and a mechanism-level test before treating the SOTA claim as settled. The signed update in Eq. (9) is likely a typo but must be fixed. The paper fits the conference scope and the ideas are worth pursuing after these additions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid, incremental UDA person search paper. It proposes two modules—PDAF, an adaptive threshold filter on backbone features, and CPR, cluster-level memory updates—and reports state-of-the-art numbers on CUHK-SYSU and PRW with ablations for each. The novelty is the specific combination and the specific filter function (Eq. 6); the ingredients have precursors, and the authors cite them.\n\nWhat the paper does well: the ablations in Tables 2 and 3 are the right experiments, and they show each component matters. CPR's one-proxy-per-cluster update is a sensible efficiency improvement, and Figure 6 supports the memory-saving claim. The code is available. For anyone working in UDA person search, this is a useful baseline and a clear reference point.\n\nNow the soft spots, in proportion. The gain on CUHK-SYSU over the runner-up is under one mAP point, and all results are single-seed with no error bars. Hyperparameters and epoch counts are tuned per target dataset, which is common in this area but limits how much weight one number can carry. Table 2 is confusingly formatted, and there are typos like \"34,7\" and \"DCSA\" vs \"DSCA.\" None of these are fatal, but they should be cleaned up.\n\nThe stress-test concern about PDAF is fair but not a rejection trigger. The paper claims the channel-magnitude threshold separates foreground from background, yet only supports this with end-to-end ablations and LayerCAM visuals. It does not measure how many true-positive foreground activations get zeroed, or compare suppression rates across domains. So the mechanism story is under-evidenced. However, the ablations do show that removing PDAF hurts across both datasets, which gives the component empirical support independent of the exact mechanism. A reviewer should ask for a feature-level analysis, not reject the paper.\n\nThe citation pattern looks fine, and there is no circular reasoning. The math is simple and correct. Bottom line: this is a competent, honest paper for a niche audience. It deserves a serious referee; with multi-seed results and a clearer ablation table it could be accepted. I'd bring it to a reading group only if someone is tracking UDA person search baselines, and I would not cite it in my own work outside that area.","headline":"Solid incremental UDA person search with useful ablations; the filter mechanism is plausible but not directly proven, and single-seed results keep the SOTA claim modest.","tokens_in":12885,"tokens_out":3060,"would_cite":false,"duration_ms":26730,"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":"Cleaning pseudo-label noise at both image and instance levels pushes unsupervised domain-adaptive person search to state-of-the-art, rivaling some fully supervised models.","keywords":["unsupervised domain adaptation","person search","pseudo-label denoising","adaptive feature filtering","cluster proxy memory","contrastive learning","pedestrian detection","re-identification"],"falsifier":"Run the trained model on the target-domain test set and, for every ground-truth person box, compute the fraction of feature channels that the Self-Calibrating Filter sets to zero; if a substantial fraction of channels on correctly detected people are zeroed out, the filter is removing foreground signal rather than only noise, contradicting the paper's explanation of its mechanism.","tokens_in":11903,"feed_emoji":"🎯","tokens_out":8105,"duration_ms":63991,"temperature":0.7,"pith_summary":"This paper tries to establish that unsupervised domain-adaptive person search can be made substantially more accurate by treating the two sources of pseudo-label noise separately: low-quality pseudo bounding boxes from the detector and misassigned pseudo identities from clustering. The proposed framework filters noisy boxes out of image-level features with an adaptive threshold in the backbone, and it cleans instance-level noise by replacing per-instance memory entries with one proxy representation per cluster. On the CUHK-SYSU and PRW benchmarks the authors report state-of-the-art results among UDA methods, with margins that make the approach competitive with some fully and weakly supervised systems. The reason this matters is that a UDA person search model that no longer drowns in its own noisy labels is closer to being usable on new camera networks without manual annotation.","feed_headline":"Dual self-calibration lifts unsupervised person search to SOTA","feed_subtitle":"Image-level filtering and cluster-level memory cut pseudo-label noise, beating prior UDA and some supervised methods.","key_machinery":"Two mechanisms carry the argument. First, the Perception-Driven Threshold derives a dynamic channel-wise cutoff from average pooling scaled by a sigmoid of an MLP on max pooling, and the Self-Calibrating Filter applies this cutoff through Eq. 6, a learnable blend of the classical soft threshold and a high-order soft threshold that approaches $(x^2 - \\tau^2)^{1/2}$. This filter block is inserted after each of the first four ResNet-50 layers to suppress background and noisy pseudo-box features. Second, the Cluster Proxy Representation keeps one proxy $c_k$ per identity cluster in a memory dictionary, initializes each proxy as the centroid of its cluster, updates it online by $c_k \\leftarrow (1-\\gamma)c_k + \\gamma f$, and re-clusters once per epoch with an EMA-matched offline reinitialization. The ablations show that removing either mechanism lowers mAP and top-1 on both datasets, which is why the paper attributes the gain to these two components.","core_discovery":"The central claim, stated on the paper's own terms, is that noisy pseudo-labels in unsupervised domain-adaptive person search are best handled by a dual self-calibration: filter unreliable image-level features in the backbone, and de-pollute the cluster memory used for identity learning. The Perception-Driven Adaptive Filter computes a per-channel threshold $\\tau = \\alpha \\odot \\operatorname{AvgPool}(F)$ with $\\alpha = \\sigma(\\operatorname{MLP}(\\operatorname{MaxPool}(F)))$, then applies a self-calibrating high-order soft threshold that interpolates from a soft to a hard filter over training. The Cluster Proxy Representation stores a single momentum-updated proxy per cluster instead of a memory slot per instance, clusters the target domain only once per epoch, and reinitializes proxies offline with exponential moving average smoothing. According to the reported experiments, these modules together reach 80.2% mAP and 81.7% top-1 on CUHK-SYSU and 39.9% mAP and 81.6% top-1 on PRW, surpassing prior UDA methods and matching or beating several weakly and fully supervised baselines.","pith_inferences":["One could test whether the Perception-Driven Threshold transfers to other detection-and-retrieval tasks, such as vehicle search, since the threshold depends only on channel statistics rather than person-specific priors.","A natural extension is to make the cluster proxy update fully online, removing the once-per-epoch re-clustering step; the risk is that drifting proxies would reintroduce the misidentification pollution CPR is designed to erase.","A stress test that artificially corrupts a known fraction of pseudo-boxes, holding everything else fixed, would isolate how much of PDAF's benefit comes from filtering exactly this type of noise."],"forward_implications":["If noisy pseudo-labels are the main bottleneck, further UDA gains should come from better denoising of boxes and identities rather than from more elaborate domain alignment alone.","Because the memory dictionary holds one proxy per cluster rather than one entry per instance, training cost and memory usage scale with the number of identities, not the number of detected boxes.","The filter is inserted inside the backbone, so the same PDAF module can be added to other end-to-end person-search architectures without changing detection or ReID heads.","Reaching parity with weakly supervised and some fully supervised baselines on these benchmarks means annotation-free person search may be practical for camera networks with modest domain shift."],"supporting_citations":[{"why":"Provides the domain-aligned person search baseline and the end-to-end architecture that DSCA extends.","marker":"Li et al. 2022"},{"why":"A prior UDA person search method that DSCA compares against and outperforms.","marker":"Almansoori, Fiaz, and Cholakkal 2024"},{"why":"Supplies the self-paced contrastive learning with hybrid memory that CPR reformulates into cluster-level proxies.","marker":"Ge et al. 2020"},{"why":"Provides the DBSCAN clustering algorithm used to assign pseudo-labels to target-domain proposals.","marker":"Ester et al. 1996"},{"why":"Source of the CUHK-SYSU dataset and the OIM memory dictionary idea that CPR moves from instance level to cluster level.","marker":"Xiao et al. 2017"},{"why":"Defines the NAE head used for joint detection and re-identification in the framework.","marker":"Chen et al. 2020b"},{"why":"Supplies the adaptive mean threshold module used as the PDT comparison baseline in ablations.","marker":"Zhao et al. 2019"},{"why":"Provides the ResNet-50 backbone into which the adaptive filter is inserted.","marker":"He et al. 2016"}],"fun_headline_variants":["Dual self-calibration silences noisy pseudo-labels in person search","Filter and cluster memory wipe pseudo-label noise in UDA person search","Dual self-calibration: cutting pseudo-label noise in person search","UDA person search: self-calibrating filter and cluster proxy hit SOTA","Self-calibrating filter + cluster proxy: UDA person search's new best"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a channel's average activation, scaled by a factor learned from its max activation, separates foreground person features from background and noisy-box features in every backbone layer and across both domains; if that separation fails after domain shift, the filter would discard useful identity information while the end-to-end ablations would not reveal where the loss occurs.","fun_headline_variants_meta":{"raw":{"variants":["Dual self-calibration silences noisy pseudo-labels in person search","Filter and cluster memory wipe pseudo-label noise in UDA person search","Dual self-calibration: cutting pseudo-label noise in person search","UDA person search: self-calibrating filter and cluster proxy hit SOTA","Self-calibrating filter + cluster proxy: UDA person search's new best"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00136,"raw_usage":{"total_tokens":5592,"prompt_tokens":1090,"completion_tokens":4502,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":706,"completion_tokens_details":{"reasoning_tokens":4404}},"tokens_in":706,"tokens_out":4502,"duration_ms":29141,"temperature":1.0,"reasoning_tokens":4404,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:30:50.291795+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained model on the target-domain test set and, for every ground-truth person box, compute the fraction of feature channels that the Self-Calibrating Filter sets to zero; if a substantial fraction of channels on correctly detected people are zeroed out, the filter is removing foreground signal rather than only noise, contradicting the paper's explanation of its mechanism.","supporting_citations":[{"cited_title":"K.; Fiaz, M.; and Cholakkal, H","cited_arxiv_id":null,"evidence_quote":"A prior UDA person search method that DSCA compares against and outperforms."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the CUHK-SYSU dataset and the OIM memory dictionary idea that CPR moves from instance level to cluster level."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the adaptive mean threshold module used as the PDT comparison baseline in ablations."}],"review_version":1}