{"id":"1960d460-ec47-4208-80dc-64758ffba95e","arxiv_id":"2412.16233","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"DPWiT, a dual pyramid network with frequency-aware encoders, reaches 74.5 average mAP on a new WiFi CSI temporal activity detection dataset, beating the best baseline by about 16 points.","lead":"This paper presents DPWiT, a neural network that detects human activities like walking, running, and falling in long, continuous WiFi signals, along with the start and end times of each activity. The authors also introduce a new dataset of 553 untrimmed WiFi recordings and report a large improvement over existing video-based detection methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim depends on an unverifiable baseline adapter: replacing every baseline's backbone with 'the same encoder used in our network' may give DPWiT its own feature extractor, and the broken code link prevents auditing.","rationale":"The paper proposes a coherent architecture and the ablation study (Table 4) internally supports the design choices: removing either branch, either encoder, max-min pooling, pyramid fusion, or SMA all degrade mAP, and the LSRE-only run at 24.7 mAP shows that the TSSE is doing the heavy lifting. Those ablations make the method plausibly effective on this dataset. What is not established is whether DPWiT is genuinely better than well-implemented vision TAD baselines, because the baseline adapter is underspecified and unauditable. The phrase 'replaced their backbones with the same encoder used in our network' is the exact load-bearing assumption; it is also the same weakness the reader identified. I do not accuse the authors of manipulating results: the margin could be real. But the burden is on the paper to release the code and define the shared encoder. The malformed code link and the 'bend' class in Fig. 3 are additional warning signs that the experimental record needs correction or clarification before the SOTA claim is accepted. Because this is a verifiability problem rather than a proven mathematical error, the reader's CONDITIONAL verdict remains appropriate until the authors supply code, equal-protocol baselines, and repeated-run statistics.","tokens_in":12145,"tokens_out":5467,"duration_ms":46955,"concrete_test":"Access the corrected repository (AVC2-UESTC/WiFiTAD); verify which layers constitute the shared encoder in the baseline adapter. Then rerun Table 3 twice for at least DyFADet, ActionFormer, and TemporalMaxer: once with the exact shared-encoder adapter as described, and once with each baseline's original temporal encoder adapted only by a fixed 1D stem for 30-subcarrier CSI input, using the same training epochs and hyperparameters. If the margin between DPWiT and the best baseline shrinks by more than ~5 mAPavg, or if the shared encoder includes TSSE/LSRE modules, the SOTA claim is not established. Also run each model over 5 seeds and report mean±std, and check whether the released labels contain 'bend'; if not, Fig. 3 is invalid as evidence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the large margin in Table 3 (74.5 vs. 58.0 mAPavg over DyFADet). The Experimental Setups state: 'we modified their pipelines to fit signal data and replaced their backbones with the same encoder used in our network.' This sentence is the fulcrum: if the shared 'encoder' is only the three CGR input layers, then the comparison isolates TAD heads and is fair; if it is the full DPTCM (TSSE with Signed Mask-Attention, LSRE, and Cross-attention Pyramid Fusion), then every baseline is already a DPWiT feature extractor and the comparison measures only prediction heads, making the SOTA claim an artifact of the adapter. The paper never defines what is shared, and the code URL is malformed ('https://github.com/A VC2-UESTC/WiFiTAD'), so the protocol cannot be checked. The qualitative support is also internally inconsistent: Fig. 3 repeatedly labels 'bend' as a GT/predicted activity, but the dataset contains only walk, run, jump, wave, fall, sit, and stand, and 'bend' is absent from Table 2. Finally, no repeated-run statistics are reported, and the test set is a random 7:3 split of one room with three volunteers, so the reported gap is not shown to be outside run-to-run noise.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DPWiT, a dual-pyramid network for temporal activity detection (TAD) in untrimmed WiFi CSI streams. The architecture uses a Temporal Signal Semantic Encoder (TSSE) with a Signed Mask-Attention mechanism and a Conv-Pool branch, a learning-free Local Sensitive Response Encoder (LSRE), and a cross-attention pyramid fusion. The authors also present a self-collected dataset of 553 WiFi CSI samples with 2,114 annotated activity segments over seven classes. They report that DPWiT achieves 74.5 mAPavg versus 58.0 for the strongest adapted vision baseline (DyFADet) while using fewer GFLOPs. Ablations support the main design choices.","tokens_in":12422,"tokens_out":4952,"duration_ms":43658,"significance":"If the reported results are reproducible, the paper makes a useful contribution: it formulates WiFi-based TAD as a distinct task, provides a new dataset for it, and proposes an efficient architecture whose frequency-decomposition hypothesis is motivated by preliminary experiments. The ablations in Table 4 give plausible evidence for the value of the main components. However, the headline state-of-the-art claim currently rests on an underspecified baseline adaptation protocol, a single unreleased dataset, and no variance estimates; the code URL is malformed and the supplementary material is not available. These issues prevent an independent verification of the empirical contribution, so the significance of the paper is substantially reduced as submitted.","major_comments":[{"comment":"The baseline comparison protocol is underspecified. The sentence 'To ensure fair comparison, we modified their pipelines to fit signal data and replaced their backbones with the same encoder used in our network' does not state which modules constitute the shared 'encoder.' If the shared encoder is only the three CGR projection layers, the comparison is reasonable; if it includes the full Dual Pyramid Temporal Context Modeling module (TSSE, LSRE, and Cross-attention Pyramid Fusion), then every baseline already uses DPWiT's feature extractor and Table 3 compares only prediction heads. The code URL 'https://github.com/A VC2-UESTC/WiFiTAD' contains a space and is not resolvable, so the protocol cannot be audited. Please define precisely which parts of DPWiT are shared with the baselines and provide either working code or a detailed layer-by-layer adapter description.","section":"Experimental Setups"},{"comment":"The evaluation has no repeated-run statistics. The whole comparison in Table 3 is performed on one self-collected dataset from a single room with three volunteers, split once with a 7:3 ratio, and no standard deviations or multiple seeds are reported. The gap of 74.5 versus 58.0 mAPavg over DyFADet could be influenced by split luck or run-to-run noise, and the same concern applies to the ablation results in Table 4. The text says 'Cross-person evaluation is provided in the supplementary material,' but no supplementary material is included in this submission. Please report means and standard deviations over multiple runs/splits, and make the dataset, the exact split, and the evaluation code available or provide results on a public benchmark.","section":"Dataset and Main Results"},{"comment":"Figure 3 internally contradicts the dataset definition. The figure repeatedly labels the activity 'bend' in the ground-truth interval and in predictions (e.g., 'bend (0.97)', 'bend (0.51)'), but Table 2 lists only walk, run, jump, wave, fall, sit, and stand, and 'bend' appears nowhere in the dataset statistics. This inconsistency undermines the qualitative evidence in the Visual Analyze paragraph. Please regenerate the figure with the correct class vocabulary or explicitly clarify whether 'bend' is part of the dataset and why it is absent from Table 2.","section":"Visual Analyze (Fig. 3)"}],"minor_comments":[{"comment":"The notation in Eq. (2) is ambiguous: the 1-norm of Q+K is not defined (row-wise versus element-wise), and the shape of W_H is said to be 'the same dimensions as Q,' which is inconsistent with the transpose W_H^T. Please clarify the exact tensor operations.","section":"Method, Eq. (2)"},{"comment":"The implementation details state that the model uses '8 TSSE and LSRE backbones as feature encoders' and that output features from the last 4 layers are used for detection, but the architecture figure and Eq. (9) describe an unspecified number L of pyramid levels. Please state explicitly that L=8 and which levels are used.","section":"Implementation Details"},{"comment":"The rows 'Baseline-ResNet1d' and 'Baseline-THAT' are not tied to the description of the two sliding-window baseline methods in Experimental Setups. Please clarify whether these are the sliding-window HAR classifiers and why they are included in the mAP comparison.","section":"Table 3"},{"comment":"There are several typographical errors that should be corrected: 'Huam Activity Analysis' in the Related Work heading, 'sematic' in Figure 2, 'TarTR' in Figure 3, 'form different methods' in the Figure 3 caption, 'ration' in Eq. (12), and 'Institude' in the author affiliation.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper has a plausible architecture and a useful new dataset, but the empirical claims are not verifiable as submitted. The main risk is the baseline adapter: if the shared encoder includes DPTCM, the SOTA claim is essentially constructed. I would ask the editor to make code and data release a condition of a revised version and to have the authors specify exactly what is shared with baselines. The malformed code URL should also be fixed before any acceptance decision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper opens a genuinely new task—temporal activity detection on untrimmed WiFi CSI—and ships a new dataset of 553 samples with 2,114 annotated segments. That alone is useful for the wireless-sensing community. The frequency-decomposition story is also reasonable: transformer branches capture low-frequency semantics, conv-pool branches capture high-frequency boundaries, and the ablations in Table 4 support each design choice. The gains from max-min pooling in LSRE and from Signed Mask-Attention are concrete and non-obvious. The writing is clear and the method is reproducible in principle, though not in practice.\n\nThe soft spots are real and load-bearing. The biggest one: the baseline comparison replaces every baseline's backbone with \"the same encoder used in our network,\" but the paper never says what that encoder is. If it is only the three CGR input layers, the comparison is fair. If it is the full DPTCM, then every baseline is already DPWiT minus the head, and the 16-point gap is trivial. The ambiguity is compounded by a malformed code URL, a dataset that is not released, and no repeated-run statistics. With three volunteers and one room, the reported margin could be within run-to-run noise; the paper gives no error bars. The qualitative figure also labels a \"bend\" activity that does not exist in the class list, which suggests either a labeling error in the figure or an undisclosed class in the evaluation. Any of these alone would be a minor issue; together they make the central claim unverifiable.\n\nThat said, the method itself is sensible and the ablations are honest. The paper does not appear to be circular or knowingly misleading; it is just under-specified at a critical point. The task and dataset are worth having, and the architecture is a reasonable starting point for future work.\n\nThis paper deserves a serious referee, but it should not be accepted without major revisions. The authors need to release the dataset and working code, clarify exactly what was replaced in each baseline, and provide repeated-run means and variances. If they do that, the paper could be a solid benchmark contribution. If not, the SOTA claim should be downgraded to \"competitive on a private dataset.\"","headline":"New WiFi-CSI temporal activity detection benchmark with a plausible dual-pyramid design and supportive ablations, but the SOTA claim is unverifiable as reported due to an underspecified baseline adapter, an unreleased dataset, and a broken code link.","tokens_in":12982,"tokens_out":1844,"would_cite":false,"duration_ms":17550,"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":"A dual-pyramid network that splits WiFi channel state information into high- and low-frequency views, adds learning-free fluctuation features, and fuses them with cross-attention claims to set a new bar for detecting daily activities in…","keywords":["WiFi CSI","temporal activity detection","dual pyramid network","frequency decomposition","Signed Mask-Attention","activity localization","wireless sensing"],"falsifier":"Run the released code with each baseline using its original backbone instead of the shared DPWiT encoder, or evaluate on a multi-room, multi-occupant WiFi CSI dataset; if the 16.5-point average-mAP gap over DyFADet shrinks substantially or reverses, the claim that DPWiT's design is responsible for the improvement would be refuted.","tokens_in":11926,"feed_emoji":"📡","tokens_out":8020,"duration_ms":58703,"temperature":0.7,"pith_summary":"WiFi signals change when people move, but turning those changes into a list of 'who did what, when' in a continuous recording is harder than classifying pre-cut clips. This paper argues that the two sub-problems—where an activity starts and ends, and which activity it is—are best served by different frequency bands: high-frequency signal components mark boundaries, low-frequency components carry category identity. The authors build DPWiT, a dual-pyramid network that learns those two frequency views separately, adds a learning-free encoder for local signal fluctuations, and fuses the two feature pyramids with cross-attention. On a new dataset of 553 untrimmed WiFi recordings containing 2,114 activity instances from seven daily activities, the method reports an average mAP of 74.5% across tIoU thresholds 0.3–0.7, versus 58.0% for the strongest adapted vision baseline. The dataset itself is another contribution, since no public WiFi temporal activity detection benchmark existed.","feed_headline":"Dual frequency pyramid lifts WiFi CSI detection to 74.5 mAP","feed_subtitle":"High- and low-frequency splits plus fluctuation features time and label seven activities in untrimmed WiFi streams.","key_machinery":"The central machinery is the Dual Pyramid Temporal Context Modeling (DPTCM) module. It contains a Temporal Signal Semantic Encoder (TSSE), which splits learning into a transformer-based branch for low-frequency semantic information and a Conv-Pool branch for high-frequency boundary details, fused by ContraNorm; a Local Sensitive Response Encoder (LSRE), which slides a window across the temporal axis and computes max-minus-min fluctuations with no learned parameters; and a Cross-attention Pyramid Fusion that bidirectionally aligns and merges the two pyramids at each scale. This machinery makes the network explicitly model two complementary views of the signal at multiple temporal scales, so that the localization and classification sub-tasks each get the frequency content that the paper argues they need.","core_discovery":"On its own terms, the paper's central discovery is that temporal activity detection in WiFi CSI is not served well by treating the signal as a generic sequence; a frequency-aware decomposition is the key inductive bias. The authors show preliminary evidence that transformer-style global attention is good at classifying activities and convolutional-pooling branches are better at localizing boundaries, and that this split is confirmed by feeding low- and high-frequency filtered signals separately. Their network, DPWiT, operationalizes this by learning two feature pyramids—a Temporal Signal Semantic Encoder with a transformer branch (using a new Signed Mask-Attention) and a Conv-Pool branch, fused by ContraNorm, and a Local Sensitive Response Encoder that computes max-minus-min in sliding windows without learned parameters. A cross-attention pyramid fusion then combines the two pyramids level by level. With this design and a new 553-sample, 2,114-instance dataset, the paper reports state-of-the-art results on the task, with the largest measured gain coming from the frequency decomposition and the Signed Mask-Attention.","pith_inferences":["If the authors' frequency-split hypothesis generalizes, the same high/low frequency decomposition may transfer to other motion-sensing modalities such as mmWave radar or UWB, whose signals also encode movement as multipath variation.","The Signed Mask-Attention and the max-min fluctuation encoder are not specific to WiFi; they could be applied to other long, noisy 1D time-series detection tasks, such as seismic event detection or ECG monitoring.","The dataset's single-room, three-volunteer collection is the main unverified boundary; a multi-environment, multi-person benchmark would test whether the 74.5 mAP margin survives domain shift.","Because the authors replace all baselines' backbones with their own encoder, an independent re-baselining with the original backbones would clarify whether the improvement comes from the encoder itself or from the dual-pyramid head."],"forward_implications":["If the claims hold, WiFi-only sensors can produce activity timelines (start, end, category) for long, untrimmed monitoring without cameras or wearables, which matters for privacy-sensitive care settings.","The frequency-decomposition finding gives a concrete design rule: use high-frequency features for boundary regression and low-frequency features for classification in temporal detection of non-visual signals.","The learning-free LSRE shows that a cheap handcrafted fluctuation feature can carry much of the localization signal, reducing compute (44.1 GFLOPs vs 304.0 for DyFADet) and the risk of overfitting.","The new dataset and benchmark let future WiFi temporal activity detection work compare against a common set of seven activities, 2,114 instances, and the tIoU 0.3–0.7 protocol."],"supporting_citations":[{"why":"Supplies the prediction-head objective (focal loss plus DIoU loss) and is the ActionFormer baseline that DPWiT beats.","marker":"Zhang, Wu, and Li 2022"},{"why":"DyFADet is the strongest adapted vision baseline; the paper measures its improvement against this method.","marker":"Yang et al. 2024"},{"why":"TemporalMaxer is a baseline whose max-pooling temporal context is contrasted with the paper's frequency-split design.","marker":"Tang, Kim, and Sohn 2023"},{"why":"TADTR is a deformable-attention temporal detection baseline adapted to CSI data.","marker":"Liu et al. 2022"},{"why":"Tridet is a baseline using relative boundary modeling for temporal action detection.","marker":"Shi et al. 2023"},{"why":"Frequency-domain analysis of self-attention as a low-pass filter motivates the TSSE low/high-frequency split.","marker":"Wang et al. 2022"},{"why":"High-frequency prior for image super-resolution is cited as vision-community validation of the frequency perspective.","marker":"Li et al. 2023"},{"why":"ContraNorm is the module used to fuse the transformer and Conv-Pool branches of TSSE.","marker":"Guo et al. 2023"},{"why":"Soft-NMS is used at inference to de-duplicate overlapping candidate segments.","marker":"Bodla et al. 2017"},{"why":"DIoU loss is the localization loss in the training objective.","marker":"Zheng et al. 2020"}],"fun_headline_variants":["WiFi activity timing via dual pyramid frequency split","Signed Mask-Attention sharpens WiFi CSI activity detection","Dual pyramid network for untrimmed WiFi activity detection","Frequency-aware WiFi sensing yields 74.5 mAP on activities","WiFi CSI: high-low frequency pyramid times seven activities"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the comparison is fair and the data are representative: all baselines were given the same encoder that is part of DPWiT, and the single-room, three-volunteer dataset is treated as an adequate testbed for the reported 74.5 mAP.","fun_headline_variants_meta":{"raw":{"variants":["WiFi activity timing via dual pyramid frequency split","Signed Mask-Attention sharpens WiFi CSI activity detection","Dual pyramid network for untrimmed WiFi activity detection","Frequency-aware WiFi sensing yields 74.5 mAP on activities","WiFi CSI: high-low frequency pyramid times seven activities"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000224,"raw_usage":{"total_tokens":1423,"prompt_tokens":868,"completion_tokens":555,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":474}},"tokens_in":484,"tokens_out":555,"duration_ms":5415,"temperature":1.0,"reasoning_tokens":474,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:11:14.206263+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released code with each baseline using its original backbone instead of the shared DPWiT encoder, or evaluate on a multi-room, multi-occupant WiFi CSI dataset; if the 16.5-point average-mAP gap over DyFADet shrinks substantially or reverses, the claim that DPWiT's design is responsible for the improvement would be refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DyFADet is the strongest adapted vision baseline; the paper measures its improvement against this method."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Frequency-domain analysis of self-attention as a low-pass filter motivates the TSSE low/high-frequency split."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ContraNorm is the module used to fuse the transformer and Conv-Pool branches of TSSE."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DIoU loss is the localization loss in the training objective."}],"review_version":1}