{"id":"05db2a1e-a3e6-4ddb-bbe7-25dfa5b73f55","arxiv_id":"2502.08155","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DGSense combines VAE-based virtual data generation with episodic training to generalize wireless sensing to unseen users, rooms, and locations without target-domain data.","lead":"This paper introduces DGSense, a domain generalization framework that uses a VAE-based virtual data generator and episodic training to let wireless sensing models recognize gestures, activities, and falls for new users, rooms, and locations without any target-domain data. If it works as claimed, it could remove a major obstacle to deploying WiFi, mmWave, and acoustic sensing systems in new environments.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unspecified and unablated ResNet pretraining is a hidden confound; the claimed cross-domain gains may stem from ImageNet transfer rather than from DGSense's virtual-data and episodic-training mechanism.","rationale":"The reader's weakest assumption correctly identifies an unexamined dependency. I agree this is the single most load-bearing issue because the feature extractor is a fixed component in all experiments and baselines, so any hidden advantage from ImageNet transfer changes the interpretation of the central claim. The paper's own text is inconsistent: 'pre-trained' appears only in the abstract and introduction, while the design sections call it simply ResNet18; this ambiguity makes the reported numbers impossible to reproduce. Other issues (unfair DA baselines, missing error bars) are real but less central: they affect the comparative claims, not the absolute feasibility of target-free generalization. A single controlled re-initialization experiment would settle whether the pretraining concern lands. Therefore I recommend keeping the reader's CONDITIONAL verdict (UNCHANGED).","tokens_in":21285,"tokens_out":13128,"duration_ms":110055,"concrete_test":"Implement DGSense as described, then run the WiFi, mmWave, and acoustic cross-domain evaluations under three conditions: (1) ResNet18 initialized from scratch (random weights); (2) ResNet18 initialized with the published ImageNet weights and fine-tuned exactly as in the paper's pipeline; (3) ResNet18 with a non-ImageNet initialization to test sensitivity. Report unseen-domain accuracy for each condition with at least 5 seeds and standard deviations. If conditions (2) and (1) differ by more than about 2 percentage points, the undisclosed pretraining is a load-bearing confound; if they are within noise, the pretraining concern is dismissed but must still be documented for reproducibility.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim is that DGSense's virtual data generation and episodic training produce domain-independent features that generalize to unseen users, rooms, and locations without target data. But every one of the three systems uses a ResNet18 feature extractor, and the abstract/intro describe it as 'pre-trained,' while the body (Sec. III-E, V-B, VI-B) never states the pretraining source, the initialization procedure, whether the backbone is frozen, or any ablation of this choice. The inputs—CSI phase images, Doppler spectrograms, compressed range-Doppler maps—are not natural images, so the presumed ImageNet weights are not obviously transferable. Since this component is present in all reported results and in the w/o-DG baseline, the large gains attributed to virtual data generation and episodic training could be confounded by the pretrained initialization: if ImageNet features already provide a good general representation for these spectrogram-like inputs, the DG modules may contribute far less than claimed; if the pretrained features are mismatched, the reported numbers may depend on an undisclosed trick. The inconsistency between 'pre-trained' in the abstract and the unqualified 'ResNet18' in the system sections makes the exact setup unverifiable. No code or weights are released, so the reader cannot check which variant produced Tables/Figs. 7-14 and the abstract's figures of 83.3%, 81.7%, 95.4%, and 95.0%.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DGSense, a domain generalization framework for wireless sensing. The framework has three main components: a VAE-based virtual data generator (single-modal or cross-modal) that augments the source-domain training set, a spatial-temporal feature extractor (ResNet18 with CBAM for images and 1DCNN for time series), and a training procedure in which a main network is trained alongside per-domain networks via losses that encourage the main feature extractor and main classifier to be compatible with all domain-specific components. The authors evaluate DGSense on WiFi gesture recognition, mmWave activity recognition, and acoustic fall detection, reporting average accuracies of 83.3% for new WiFi users, 81.7% for new WiFi rooms, 95.4% for new mmWave users and locations, and 95.0% for new acoustic users, with comparisons against no-DG baselines, domain adaptation methods, and prior wireless sensing systems. The central claim is that once the model is built, it generalizes to unseen domains without any target-domain data.","tokens_in":21567,"tokens_out":9502,"duration_ms":79078,"significance":"If the reported results hold, the paper would be a useful step toward practical wireless sensing because it addresses the domain-shift problem without target-domain data and demonstrates the approach across three different wireless modalities on commodity hardware. The evaluation is genuinely broad: three sensing tasks, different signal types, multiple leave-one-domain-out protocols, and ablations of the generator type, generation method, generalization method, and number of source domains/samples. The paper also explicitly acknowledges limitations such as single-person scenarios and predefined gesture vocabularies in Section VIII. However, the empirical claims currently rest on several unspecified design choices and on comparisons that are not always aligned with the baselines' intended settings. The main strength is the breadth of real-world evaluation; the main weakness is that the attribution of the gains to the proposed mechanism is not yet fully supported by the reported experiments.","major_comments":[{"comment":"The abstract and the framework description call the spatial feature extractor a 'pre-trained Residual Network (ResNet)', and Sections V-B and VI-B say the feature extractor is 'ResNet18' (with CBAM in Section V-B), but the manuscript never states the pretraining source, the initialization procedure, whether the backbone is frozen or fine-tuned, or any ablation of this choice. Since this backbone is used in every reported system and in the w/o DG baseline, the cross-domain improvements attributed to virtual data generation and episodic training cannot be separated from the effect of the pretrained initialization. Please specify the pretraining setup and add an ablation (e.g., random initialization versus the chosen pretrained weights) with both in-domain and cross-domain accuracy; this is needed for reproducibility and for the attribution claim.","section":"Abstract, Secs. III-E, V-B, VI-B"},{"comment":"The comparison protocol is inconsistent with the design of the baselines. The text states that OneFi requires one or few labeled target samples (Sec. II-A1) and CsiGAN uses semi-supervised GAN training with some target-domain data (Sec. II-A1), yet Fig. 8 is obtained by letting 'each method utilized only the source domain data to train the model' (Sec. IV-C6). This removes the very information those methods are designed to exploit, so the conclusion that 'our method consistently achieved the highest accuracy' in this figure does not support superiority over OneFi or CsiGAN in their intended settings. Please either compare in the settings those methods define, or clearly report the comparison as a zero-shot constraint that lies outside the baselines' design, and include the numerical w/o DG values underlying Fig. 8.","section":"Sec. IV-C6"},{"comment":"The procedure described is not the episodic training of Li et al. [34], contrary to the statement that the paper adopts their strategy. The cited method forms meta-train/meta-test episodes from the source domains and updates the model on a pseudo-test domain to simulate domain shift. Here, for each source domain i, losses (13)-(15) are all computed on samples from that same domain with no held-out domain; the domain networks and the main network are trained on the same data, and the main network is never evaluated against a domain that was excluded from its parameter update during training. At minimum, the name should be changed to something like a multi-domain feature-critic training procedure, or the algorithm should be extended with actual meta-test episodes; as written, the paper does not provide evidence that 'episodic training' in the literature's sense is the mechanism behind the reported gains.","section":"Sec. III-E3, Eqs. (13)-(16)"},{"comment":"The framework has at least five free parameters — λ in Eqs. (5) and (7), ω1 and ω2 in Eqs. (6) and (9), θ1 and θ2 in Eq. (16), and the virtual-to-real sample ratio — but no values or sensitivity analysis are reported anywhere, and no code or trained models are released. Given that Tables IV-VI and Figs. 7-14 are the entire empirical support for the central generalization claim, the absence of these settings makes the results unverifiable and prevents a reader from assessing whether the reported margins are robust. Please add a hyperparameter table and per-fold or leave-one-domain-out results with variance, and consider releasing the code.","section":"Secs. III-D, III-E, VII"}],"minor_comments":[{"comment":"The virtual-data quality check is presented as verifying that virtual data 'followed the same distribution' as real data, but the protocol trains on real data and tests on virtual data generated from the same real samples via a VAE trained on those samples; high accuracy in this setup mostly reflects reconstruction fidelity. Please rephrase as a reconstruction-quality check and rely on the held-out-domain results in Section VII for the claim that virtual data improve generalization.","section":"Sec. IV-C2, Sec. V-C2"},{"comment":"The symbol Ds is overloaded: Eq. (1) defines Ds as a set of source domains, while Eq. (3) redefines Ds as the training set containing all training samples. Please use distinct notation for the domain set and the training set.","section":"Sec. III-A"},{"comment":"The outlier detector mentioned in Section VIII is not described or evaluated anywhere in the paper; either remove the reference or provide details of how it is trained and how it affects the reported results.","section":"Sec. VIII"},{"comment":"All reported accuracies are point estimates without standard deviations, confidence intervals, or per-fold values. With only four to six domains per experiment, the differences in tables such as Table V (95.3 vs. 96.5 vs. 97.5) may be within noise; please report the variability across folds or leave-one-domain-out splits.","section":"Tables I-VI, Figs. 7-14"},{"comment":"The preprocessing description is underspecified for reproducibility: window sizes for the moving average or median filter, the threshold values for filtering, and the parameters of the Power Burst Curve segmentation are not given.","section":"Sec. III-C"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The paper has a solid empirical core, but the current text overclaims on two fronts: the role of the pretrained ResNet weights is unexamined, and the comparison with domain-adaptation baselines is not carried out in their intended settings. The 'episodic training' label is also inaccurate relative to the cited work. These issues are fixable in revision, so I recommend major revision rather than rejection. The authors should be encouraged to release code and hyperparameters, because the value of this paper depends heavily on reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent engineering paper, and the central claim—DGSense improves generalization to unseen users, rooms, and locations without target data—is probably true. The cross-modal VAE generator for multi-modal WiFi samples is the most original piece; episodic training is borrowed from Li et al. but applied cleanly to three sensing tasks. The evaluation is a real strength: three separate real-world data collections, consistent gains over the w/o DG baseline, and ablations over generator type, DG method, source-domain count, and real/virtual sample counts. Table VI deserves credit: GAN and MMD baselines get 25% or 50% target data while DGSense gets none, and DGSense still wins. The limitations section is also honest about unseen gestures and multi-person scenarios.\n\nSoft spots, in order of importance.\n\n1. The pretrained ResNet issue is real, but not as a confound of the relative gains. Since the w/o DG baseline uses the same feature extractor, the incremental effect of virtual data and episodic training is still isolated. The problem is that the absolute numbers and the 'domain independent features' claim cannot be checked without knowing the pretraining source, whether the backbone is frozen, and an ablation of random initialization. The abstract says 'pre-trained'; the body just says ResNet18. That is an unforced reporting gap.\n\n2. The OneFi/CsiGAN comparison in the WiFi section is unfair. Those methods are designed for one-shot or semi-supervised domain adaptation and need target data; running them with none and declaring victory is not a meaningful comparison. The authors actually acknowledge this in the text, which softens the blow, but the section should be reworked to give those baselines their intended target data, or relabeled as a zero-target-data variant.\n\n3. No error bars, no hyperparameter values, no code or data release. With free parameters like lambda, omega1/2, theta1/2, and the virtual-to-real ratio, the results are hard to verify. I do not see evidence of cherry-picking, but the paper should report these details.\n\nAlso, the abstract's claim to 'eliminate the domain dependence problem' overstates what is shown; this is an improvement, not an elimination.\n\nWho is this for? People working on wireless sensing and ubiquitous computing, especially DG/DA for WiFi, mmWave, and acoustic signals. It is not a methodological breakthrough, but it is a solid benchmark-type contribution. I would send it to peer review—it deserves serious referee time. After the baseline and pretraining issues are fixed, I would cite it; as is, I would cite it cautiously and only with those caveats in mind.","headline":"A solid engineering paper on domain generalization for wireless sensing whose empirical claim is probably right, but the unstated ResNet pretraining and an unfair DA baseline comparison need fixing before I would trust the numbers.","tokens_in":22111,"tokens_out":3441,"would_cite":true,"duration_ms":37764,"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":"DGSense reports that a wireless sensing model trained only on source domains keeps working for unseen users, rooms, and locations, with no target-domain data, across WiFi, mmWave, and acoustic signals.","keywords":["domain generalization","wireless sensing","episodic training","virtual data generation","WiFi CSI gesture recognition","mmWave activity recognition","acoustic fall detection","cross-modal variational autoencoder"],"falsifier":"Retrain the three DGSense pipelines with randomly initialized ResNet18 weights instead of pre-trained ones, holding every other component fixed; if the new-domain accuracies (83.3%, 81.7%, 95.4%, 95.0%) fall to the reported w/o-DG levels (roughly 20–25% on WiFi), then the pretrained initialization, not DGSense's domain-independence mechanism, carries the generalization result.","tokens_in":21018,"feed_emoji":"📡","tokens_out":9743,"duration_ms":73224,"temperature":0.7,"pith_summary":"This paper proposes DGSense, a domain-generalization framework for wireless sensing, and reports that a sensing model trained on a few source domains keeps working in unseen domains—new users, new rooms, new locations—without any data from those domains and without retraining. The framework is designed to be general across sensing technologies, and it is demonstrated on WiFi gesture recognition, mmWave activity recognition, and acoustic fall detection. Reported average accuracies on unseen domains are 83.3% for new WiFi users, 81.7% for new WiFi rooms, 95.4% for new mmWave users and locations, and 95.0% for new acoustic users. A reader should care because existing wireless-sensing solutions for the domain-shift problem rely on domain adaptation and still need target-domain data; DGSense targets the stronger setting where the target domain is entirely absent at training time.","feed_headline":"One training pass lets wireless sensing jump to unseen users and rooms","feed_subtitle":"New WiFi, mmWave, and acoustic users and rooms get 83–95 percent accuracy with zero target-domain data.","key_machinery":"The load-bearing object is the episodic-training loop paired with the virtual data generator. Each source domain gets its own feature extractor and classifier; a shared main network is trained by cycling through domains with three losses: the main extractor through the main classifier, the main extractor through each domain classifier, and each domain extractor through the main classifier. Forcing the main extractor to be readable by every domain classifier—and the main classifier to read every domain extractor—is what the paper identifies as the source of domain-independent features. The virtual data generator is a VAE: for images a single encoder/decoder pair, and for multi-modal WiFi data a cross-modal design in which one base modality (amplitude) is encoded and all modalities are decoded from the shared latent code, preserving inter-modal consistency. Spatial features come from ResNet18 with CBAM within each residual block, temporal features from a 1DCNN. The t-SNE plots and the rising per-episode accuracy of the held-out user (24.2% to 83.3% over five episodic steps) are presented as direct evidence that the main extractor's features become more domain-independent as episodic training proceeds.","core_discovery":"On its own terms, the paper's discovery is that the domain-dependence problem in wireless sensing can be attacked with a general, task-agnostic recipe rather than task-specific signal processing: generate diverse virtual training data with a VAE-based generator (single-modal for images, cross-modal for multi-modality WiFi samples so that amplitude, phase, and spectrogram remain consistent), then train a main feature extractor and classifier episodically against per-domain feature extractors and classifiers so that the main network must classify through every domain's eyes. At test time only the main network runs. The paper reports that this recipe yields high accuracy in unseen domains—83.3% for new WiFi users, 81.7% for new rooms, 80.6% for new users in new rooms, 95.4% for new mmWave users and locations, and 95.0% for new acoustic users—compared with roughly 20–25% on WiFi without the framework. The feature extractors combine a pre-trained ResNet18 with CBAM attention for spatial inputs and a 1DCNN for temporal inputs.","pith_inferences":["One implication the paper leaves implicit is that the claimed domain independence is only demonstrated within the range of variation spanned by the source domains (users, rooms, locations); a harder hold-out that differs in channel geometry, hardware, or frequency band would test whether the learned features are truly domain-independent rather than interpolation across familiar variations.","Because the VAE is trained on source-domain data, the virtual samples can only diversify the training set within the source distribution. A testable extension would be to monitor virtual-data quality (the reported roughly 97% agreement with real data) as the number of source domains shrinks, since the generator's diversity is bounded exactly where the framework needs it most.","The abstract's \"pre-trained\" ResNet is never ablated; if the pretrained initialization is the dominant contributor, then a much simpler sensing model might reproduce the new-domain accuracies, and the episodic-training contribution would need to be re-measured relative to that baseline."],"forward_implications":["If DGSense generalizes as reported, a wireless sensing deployment can be built from a few users and rooms and shipped to new environments without a separate data-collection trip.","The same framework works across WiFi, mmWave, and acoustic signals, so a single training recipe can serve gesture recognition, activity recognition, and fall detection.","The cross-modal generator matters: replacing it with a multi-modal generator that noisifies each modality separately drops new-user WiFi accuracy (for example, from about 84.6% to 73.8% in one test room), so preserving inter-modal consistency is part of the generalization gain.","Performance scales with source diversity: 4–5 source domains and 16–20 real samples per class suffice, so the framework is practical in data-scarce settings.","Inference is real-time (42.7 ms for WiFi, 64.1 ms for mmWave, and 849.6 ms for acoustic including preprocessing), so the generalization benefit does not come at the cost of deployability."],"supporting_citations":[{"why":"Supplies the episodic training strategy between a main network and per-domain networks that DGSense adapts to wireless sensing.","marker":"[34]"},{"why":"Supplies ResNet18, the spatial feature extractor used for phase images, Doppler spectrograms, and compressed Doppler maps.","marker":"[36]"},{"why":"Supplies the CBAM attention modules inserted in each residual block for spatial and channel attention.","marker":"[37]"},{"why":"Motivates using autoencoder-based virtual data generation to improve generalization; DGSense's VAE generator extends this idea.","marker":"[35]"},{"why":"Provides the two-way test (train on real, test on virtual, and vice versa) used to verify virtual data quality.","marker":"[29]"},{"why":"Baseline one-shot WiFi recognition method that still needs target-domain labels; comparing against it shows the value of zero-target-data generalization.","marker":"[16]"},{"why":"Baseline semi-supervised GAN method that requires target-domain data; the comparison supports the paper's claim of generalizing without such data.","marker":"[14]"},{"why":"Baseline acoustic fall-detection method and source of the Doppler-spectrogram representation used for the acoustic sensing system.","marker":"[26]"}],"fun_headline_variants":["Wireless sensing that adapts to unseen users without new data","Train once, sense anywhere: domain-general wireless sensing","Zero-shot wireless sensing: high accuracy in unseen rooms and users","New framework makes wireless sensing robust to new environments","Domain generalization for sensing: no retraining needed for new spaces"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a ResNet18 the abstract calls pre-trained gives useful starting features for inputs that are not natural images—WiFi phase data, Doppler spectrograms, and compressed range-Doppler maps—and the paper never reports the pretraining source or an ablation without it.","fun_headline_variants_meta":{"raw":{"variants":["Wireless sensing that adapts to unseen users without new data","Train once, sense anywhere: domain-general wireless sensing","Zero-shot wireless sensing: high accuracy in unseen rooms and users","New framework makes wireless sensing robust to new environments","Domain generalization for sensing: no retraining needed for new spaces"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000774,"raw_usage":{"total_tokens":3488,"prompt_tokens":1071,"completion_tokens":2417,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":687,"completion_tokens_details":{"reasoning_tokens":2350}},"tokens_in":687,"tokens_out":2417,"duration_ms":20548,"temperature":1.0,"reasoning_tokens":2350,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T10:13:05.112736+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the three DGSense pipelines with randomly initialized ResNet18 weights instead of pre-trained ones, holding every other component fixed; if the new-domain accuracies (83.3%, 81.7%, 95.4%, 95.0%) fall to the reported w/o-DG levels (roughly 20–25% on WiFi), then the pretrained initialization, not DGSense's domain-independence mechanism, carries the generalization result.","supporting_citations":[{"cited_title":"Episodic training for domain generalization,","cited_arxiv_id":null,"evidence_quote":"Supplies the episodic training strategy between a main network and per-domain networks that DGSense adapts to wireless sensing."},{"cited_title":"Deep residual learni ng for image recognition,","cited_arxiv_id":null,"evidence_quote":"Supplies ResNet18, the spatial feature extractor used for phase images, Doppler spectrograms, and compressed Doppler maps."},{"cited_title":"Cbam: Convol utional block attention module,","cited_arxiv_id":null,"evidence_quote":"Supplies the CBAM attention modules inserted in each residual block for spatial and channel attention."},{"cited_title":"Learning to learn single d omain generalization,","cited_arxiv_id":null,"evidence_quote":"Motivates using autoencoder-based virtual data generation to improve generalization; DGSense's VAE generator extends this idea."},{"cited_title":"Deep adaptati on networks based gesture recognition using commodity WiFi,","cited_arxiv_id":null,"evidence_quote":"Provides the two-way test (train on real, test on virtual, and vice versa) used to verify virtual data quality."},{"cited_title":"OneFi: One-shot reco gnition for unseen gesture via COTS WiFi,","cited_arxiv_id":null,"evidence_quote":"Baseline one-shot WiFi recognition method that still needs target-domain labels; comparing against it shows the value of zero-target-data generalization."},{"cited_title":"CsiGAN: Robust channe l state information-based activity recognition with GANs,","cited_arxiv_id":null,"evidence_quote":"Baseline semi-supervised GAN method that requires target-domain data; the comparison supports the paper's claim of generalizing without such data."},{"cited_title":"Fall detection via inaudible acoustic sensing,","cited_arxiv_id":null,"evidence_quote":"Baseline acoustic fall-detection method and source of the Doppler-spectrogram representation used for the acoustic sensing system."}],"review_version":1}