{"id":"6e93a746-233a-4229-9d70-3fd8b014feea","arxiv_id":"1908.10307","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Thermal imaging can track physiological and affective states, but most evidence comes from controlled laboratories, and this survey organizes pipelines and open challenges for moving to real-world use.","lead":"This survey reviews how thermal cameras can measure breathing, heart rate, and stress from skin temperature without physical contact. It maps the field and identifies what must improve before cheap, mobile thermal sensors can monitor health and emotions in daily life.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Survey's 99.7% k-fold accuracy claim (§3.2) contradicts its own LOSO recommendation (§4.3) and likely reflects subject leakage; TIPA toolkit is not shipped, leaving the central pipeline claim unverifiable.","rationale":"The reader's verdict is CONDITIONAL, and the stated weakest assumption concerns external validity: automatic ROI tracking and temperature handling may fail outside the tested environmental conditions. That is a legitimate limitation, but it is one the authors explicitly acknowledge in §4.2. The stronger concern is internal: the survey's own reported 99.7% accuracy is obtained with a methodology (random 10-fold CV on physiological data) that the same paper identifies as prone to artificially high results. This is not an external challenge; it is an inconsistency within the manuscript. If the 99.7% figure is inflated by subject leakage, then the survey's most striking quantitative evidence for its own pipeline—and by extension for the feasibility of ubiquitous mobile thermal imaging—is not trustworthy. The associated TIPA toolkit, promised in §4.3 and §5, is not shipped, so the pipeline cannot be verified by others. In combination, these issues confirm the reader's CONDITIONAL verdict rather than overturning it: the survey's organizational content is useful, but its self-evaluated performance claims require reproducible details and subject-independent evaluation before they can be taken as definitive. The test proposed here would settle whether the 99.7% number reflects genuine performance or a methodological artifact.","tokens_in":28740,"tokens_out":5082,"duration_ms":53246,"concrete_test":"Reproduce the §3.2 re-analysis on the released DeepBreath dataset with subject-stratified random folds: assign all temporal clips of each participant to the same fold and compute 10-fold accuracy. Compare against the reported 99.7% and against the original LOSO result. If the subject-stratified accuracy drops to the ~84.59% LOSO level, the reported 99.7% is attributable to same-subject leakage and the survey's pipeline claim loses its headline support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is internal to the survey's own evidence. In §3.2 the authors state that re-running their DeepBreath stress-recognition system on the released dataset with 10-fold cross-validation yields 99.7% accuracy, whereas the original study reported 84.59% using leave-one-subject-out (LOSO) cross-validation. Yet §4.3 of this same survey warns that on human physiological data, non-LOSO k-fold CV can yield 'artificially high results due to training and testing a machine learning model on temporally adjacent samples' and recommends LOSO as the safeguard. The 99.7% figure is reported without any details about folds, subject splitting, or code, so it cannot be checked. Given the warning in the same paper, random 10-fold CV almost certainly mixes clips from the same participant across train and test, inflating accuracy far above the subject-independent 84.59%. The survey then uses this number to argue that the respiration-variability pipeline is state-of-the-art (§3.2), i.e., as direct support for the central claim that thermal-imaging pipelines to affective states have been established. If the 99.7% figure is the product of leakage, the survey's main quantitative evidence for its own pipeline is misleading. The promised TIPA toolkit is not released, so the pipeline cannot be independently run. This is not a challenge from outside consensus; it is an internal methodological inconsistency that undermines one of the few quantitative demonstrations offered.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This survey reviews the literature on thermal imaging for physiological and affective computing, with the stated aim of establishing computational and methodological pipelines from thermal images of the skin to affective states, and with a particular emphasis on mobile, low-cost thermal cameras for real-world applications. The paper organizes the literature around four physiological thermal signatures (cardiovascular, perspiratory, respiratory, and muscular), summarizes experimental protocols and system specifications in tables, and discusses challenges such as ROI tracking, ambient temperature variation, and evaluation methodology. It also makes prescriptive claims, including a recommendation for leave-one-subject-out (LOSO) cross-validation and a critique of nonstandard metrics such as CAND.","tokens_in":29039,"tokens_out":3716,"duration_ms":36923,"significance":"If the findings are reliable, this survey provides a useful systematization of a scattered and methodologically heterogeneous literature. It performs a service by flagging nonstandard evaluation metrics, highlighting contradictory results (e.g., chin temperature responses in stress), and identifying the need for subject-independent evaluation. The emphasis on mobile thermal imaging and the promise of open-source tooling and datasets are valuable for the community. However, the paper's central claim depends on several quantitative claims from the authors' own prior work, and one of those claims—the 99.7% k-fold accuracy—directly conflicts with the paper's own methodological guidance, weakening the survey's credibility as a neutral assessment.","major_comments":[{"comment":"The reported 99.7% accuracy from a 10-fold cross-validation of the DeepBreath system is presented as evidence supporting the state-of-the-art claim, but the paper does not specify whether folds were split by participant. Given the paper's own warning in §4.3 that non-LOSO k-fold cross-validation can produce 'artificially high results due to training and testing a machine learning model on temporally adjacent samples,' this figure is likely inflated by subject leakage and is not directly comparable to the 84.59% LOSO result. The authors should either provide a subject-independent evaluation with full splitting details or remove the 99.7% figure from the claim.","section":"§3.2"},{"comment":"The survey promises the release of the TIPA open-source toolkit ('Following this review, we release an open-source toolkit for Thermal Imaging-based Physiological and Affective computing'), but the manuscript does not provide a working URL, repository, or documentation for this toolkit. Without the actual release, the central claim of establishing computational and methodological pipelines cannot be independently verified. The authors should either provide the toolkit and a stable link, or remove the promise and temper the corresponding contribution statement.","section":"§4.3 and §5"},{"comment":"The re-analysis of Hamedani et al. reports a Pearson correlation of r=0.58 between the thermal-imaging heart-rate estimates and reference PPG signals, but the aggregation procedure is not described: it is unclear whether this is a per-participant average, a pooled correlation, or computed over a specific time window. The critique of the CAND metric is well taken, but the quantitative re-analysis should be specified so that the reader can assess its validity.","section":"§2.2"}],"minor_comments":[{"comment":"The sentence 'the work reported in [74] achieved strong correlations of sequential respiratory rates with the ground truth (r=0.974)' appears to cite the tone-mapping paper by Ledda et al.; according to Table 2, the r=0.974 result is from Pereira et al. [93]. Please correct the citation.","section":"§2.4"},{"comment":"The statement 'Following this review, we release an open-source toolkit' is ambiguous about timing and availability; if the toolkit is not yet released, rephrase to avoid promising a resource that is not accessible at the time of publication.","section":"§4.3"},{"comment":"If the 99.7% accuracy is retained after subject-independent evaluation, the authors should also report the standard deviation across folds and clarify that the 10-fold and LOSO results are not directly comparable.","section":"§3.2"},{"comment":"The use of bold brackets to indicate values that are lower than the system's specifications is not explained in the table caption; a brief note would help the reader interpret the table.","section":"Table 4"},{"comment":"The phrase 'average purse rate' appears to be a typo for 'average pulse rate'; please correct it.","section":"§2.2"}],"recommendation":"major_revision","confidential_remarks":"The survey leans heavily on the authors' own prior publications for its key quantitative claims (DeepBreath, ThermSense, Nose Heat), which is not disqualifying but warrants careful checking of the 99.7% figure and the TIPA toolkit availability before acceptance. The paper's internal contradiction between §3.2 and §4.3 is the main load-bearing issue; it is fixable, but the authors should either justify the k-fold result with subject-independent details or retract it. The fit with the journal should also be confirmed given the survey's strong methodological agenda."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about arXiv:1908.10307. First, this is a genuinely useful survey: it organizes a scattered field around a pipeline (ROI selection, thermal signature, metric, affective state) and it does the field a service by re-analyzing published heart-rate results that used CAND, showing the 92.46% \"accuracy\" is only r=0.58 with the standard metric. Second, the survey undermines its own credibility when it reports 99.7% accuracy for its own stress-recognition system using k-fold CV, because §4.3 of the same paper warns that non-LOSO CV on human physiological data can produce artificially high results due to temporally adjacent samples. No details of the split are given, so the number is unverifiable and likely reflects leakage.\n\nWhat the paper does well: the review is careful. It flags contradictory findings (chin temperature stress drop in Engert vs no change in Veltman) and acknowledges that most studies don't directly evaluate the physiological signatures. The summary tables of camera specs and protocols are solid. The CAND re-analysis is a concrete, reproducible contribution (they used data in the original paper's Table 1). The self-citations are heavy, especially in the mobile sections, but the central physiological claims are grounded in many external studies, so that's not disqualifying.\n\nWhere it's soft: the 99.7% figure is load-bearing because it's used to argue the respiration-variability pipeline is state-of-the-art. If it's a leakage artifact, the survey's main quantitative support for its own pipeline collapses. The promised TIPA toolkit is not actually available at the given link, so the pipeline abstraction can't be independently exercised. The authors' admission that their mobile experiments don't cover humidity or climate extremes is a fair limitation, not a fatal flaw.\n\nVerdict: this deserves peer review. With the 99.7% number made reproducible (or removed and replaced with the LOSO result) and the toolkit shipped, it could be a standard reference. As it stands, read it for the survey, don't cite the 99.7%.","headline":"A useful survey with a real methodological service (the CAND re-analysis), but its own headline 99.7% number is reported without the subject-independent validation the paper itself demands.","tokens_in":29543,"tokens_out":2618,"would_cite":true,"duration_ms":24118,"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":"This survey makes the case that thermal images of the skin can be read as physiological signals—respiratory, cardiovascular, perspiratory, muscular—and that low-cost mobile thermal cameras can move this contactless affective sensing from…","keywords":["thermal imaging","thermography","affective computing","physiological computing","mental stress","skin temperature","respiration monitoring","mobile thermal cameras"],"falsifier":"Take a low-cost mobile thermal camera to a humid outdoor site (e.g., a seaside or near a pool) on a hot day, have a person breathe at a known pace, and compare the pipeline's estimated breathing rate against a chest belt; if the correlation drops markedly or the face region cannot be tracked, the central claim of ubiquitous mobile thermal imaging is not supported.","tokens_in":28557,"feed_emoji":"🌡️","tokens_out":6032,"duration_ms":59168,"temperature":0.7,"pith_summary":"This survey argues that thermal images of the skin can be read as physiological signals—cardiovascular, respiratory, perspiratory, and muscular—and that these signals can be linked to affective states such as stress, fear, startle, and love. Its central claim is that a pipeline of region-of-interest selection, automatic tracking, spatial interpretation, and metric extraction turns thermograms into affective measurements. The paper further claims that new low-cost, small thermal cameras make this monitoring possible outside the lab, in mobile and real-world settings. A reader should care because contactless temperature-based monitoring could serve applications from stress-aware workstations to healthcare monitoring without requiring worn sensors or good lighting.","feed_headline":"Thermal cameras can now read stress and emotion from skin heat","feed_subtitle":"Cheap phone-attached thermal sensors can take contactless physiological monitoring from the lab to real life.","key_machinery":"The load-bearing machinery is the four-stage computational pipeline proposed and reviewed by the paper: (1) region-of-interest (ROI) selection on skin, such as the nose tip, nostrils, perinasal area, or finger; (2) automatic ROI tracking, using methods such as the thermal gradient flow (TGF) tracker and 'optimal quantization' that adapts the temperature-to-image mapping against environmental temperature drift; (3) spatial interpretation, typically averaging temperatures over the ROI or integrating thermal voxels; and (4) metrics and features, from simple temperature directional change and slope to variability metrics and the respiration variability spectrogram. The key mechanism is that blood-flow regulation, sweating, breathing airflow, and muscle activity each leave distinctive temperature patterns on the skin that can be separated by choosing the right ROI and interpretation method.","core_discovery":"On the paper's own terms, the central discovery is that thermography of the human skin carries multiple physiological signatures that can be computationally linked to affective states: vasoconstriction and vasodilation change skin temperature (notably at the nose tip), sweat gland activation changes perinasal and finger temperatures, the breathing cycle changes nostril and mouth temperatures, and facial muscle contractions change local temperatures. The survey establishes a computational and methodological pipeline—ROI selection, automatic ROI tracking, spatial interpretation, and metric/feature computation—that connects raw thermal video to physiological time series and then to affective labels. It concludes that low-cost mobile thermal cameras, despite lower resolution and unstable sampling rates, are sufficient for many of these measurements, and that recent work has already demonstrated robust respiratory tracking and automatic stress recognition in unconstrained outdoor conditions.","pith_inferences":["The pipeline's logic could extend beyond the affective states reviewed: the same ROI-tracking and interpretation machinery might be reused for continuous health metrics such as fever screening, dehydration, or respiratory-rate monitoring in daily life.","If standard evaluation metrics (Pearson correlation, leave-one-subject-out cross-validation) were uniformly adopted, several optimistic accuracy reports in the literature would likely be revised downward—including the field's own cardiac-pulse estimates—refocusing research on signal quality.","Combining thermal cameras with a standard RGB camera could stabilise ROI tracking in extreme humidity or heat, but the added hardware and computation may undercut the portability that makes mobile thermal sensing attractive, so a purely thermal solution remains the key engineering target.","The privacy profile of thermal imaging—it does not capture facial identity under ordinary use—could make it the preferred unobtrusive sensing modality in sensitive settings like bedrooms, locker rooms, and hospitals, a consequence the survey implies but does not develop."],"forward_implications":["Nose-tip temperature drops have been observed across independent studies of mental stress, fear, and cognitive load, making thermal directional change a candidate non-contact stress indicator.","Respiratory rate can be extracted from nostril ROI temperatures with very high correlation to a reference belt (r=0.9987) even when the person is walking outdoors or climbing stairs using a low-cost camera.","Automatic affect recognition from thermal signatures is possible: a deep-learning system using a respiration variability spectrogram achieved 84.59% accuracy in leave-one-subject-out stress detection.","Perspiratory activity, measured from perinasal and finger regions, correlates strongly (r up to 0.968) with standard electrodermal activity sensors, offering a contactless proxy for sympathetic arousal.","Because thermal imaging is insensitive to ambient light and does not require skin contact, it is a practical alternative to remote PPG and worn sensors in dark or healthcare settings."],"supporting_citations":[{"why":"Supplies the optimal quantization and thermal gradient flow tracking methods used to make respiratory measurement robust outdoors.","marker":"[16]"},{"why":"Demonstrates perspiratory thermal signatures with high correlation to electrodermal activity.","marker":"[90]"},{"why":"Provides the stress-task evidence that nose, chin, and corrugator temperatures drop under stress.","marker":"[30]"},{"why":"Introduces the respiration variability spectrogram and deep-learning stress detection that grounds automatic affect recognition.","marker":"[12]"},{"why":"Combines nasal thermal variability with blood volume pulse for end-to-end instant stress detection.","marker":"[15]"},{"why":"Foundational cardiac pulse extraction method from thermal video that later work optimizes.","marker":"[36]"},{"why":"Introduces the pore activation index from high-resolution thermal imaging of sweat pores.","marker":"[69]"},{"why":"Early machine-learning affect classification from five facial ROIs with LDA.","marker":"[86]"},{"why":"Establishes respiration rate, inter-breath interval, and relative tidal volume extraction from thermography.","marker":"[75]"},{"why":"Shows nose temperature decreases under driving simulation stress, a key directional-change result.","marker":"[87]"}],"fun_headline_variants":["Skin heat reveals stress and emotion without contact","Cheap thermal cameras bring emotion tracking to real life","Thermal imaging reads your feelings from skin temperature","Mobile thermal sensors track stress and affect in the wild","No-contact stress detection via affordable thermal imaging"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The promise of mobile thermal imaging in everyday settings depends on the assumption that automatic region tracking and temperature handling stay accurate outside the tested conditions—the authors note that swimming pools, the seaside, humidity, extreme heat, and other climates have not been covered; if that tracking fails there, the central claim of ubiquitous mobile thermal sensing is weakened.","fun_headline_variants_meta":{"raw":{"variants":["Skin heat reveals stress and emotion without contact","Cheap thermal cameras bring emotion tracking to real life","Thermal imaging reads your feelings from skin temperature","Mobile thermal sensors track stress and affect in the wild","No-contact stress detection via affordable thermal imaging"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000256,"raw_usage":{"total_tokens":1517,"prompt_tokens":831,"completion_tokens":686,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":447,"completion_tokens_details":{"reasoning_tokens":615}},"tokens_in":447,"tokens_out":686,"duration_ms":7161,"temperature":1.0,"reasoning_tokens":615,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:46:20.500453+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a low-cost mobile thermal camera to a humid outdoor site (e.g., a seaside or near a pool) on a hot day, have a person breathe at a known pace, and compare the pipeline's estimated breathing rate against a chest belt; if the correlation drops markedly or the face region cannot be tracked, the central claim of ubiquitous mobile thermal imaging is not supported.","supporting_citations":[{"cited_title":"Julier, Nicolai Marquardt, and Nadia Bianchi-Berthouze","cited_arxiv_id":null,"evidence_quote":"Supplies the optimal quantization and thermal gradient flow tracking methods used to make respiratory measurement robust outdoors."},{"cited_title":"Pavlidis, P","cited_arxiv_id":null,"evidence_quote":"Demonstrates perspiratory thermal signatures with high correlation to electrodermal activity."},{"cited_title":"Grant, Daniela Cardone, Anita Tusche, and Tania Singer","cited_arxiv_id":null,"evidence_quote":"Provides the stress-task evidence that nose, chin, and corrugator temperatures drop under stress."},{"cited_title":"Julier, and Nadia Bianchi-Berthouze","cited_arxiv_id":null,"evidence_quote":"Combines nasal thermal variability with blood volume pulse for end-to-end instant stress detection."},{"cited_title":"Krzywicki, Gary G","cited_arxiv_id":null,"evidence_quote":"Introduces the pore activation index from high-resolution thermal imaging of sweat pores."}],"review_version":1}