{"id":"895884d4-8bda-4397-9945-f22b6b53fd2a","arxiv_id":"2507.14885","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A 14.86k-parameter spectral attention model combining Chirp-Z zoom and unsupervised contrastive learning reaches near state-of-the-art cross-dataset heart rate accuracy under motion.","lead":"BeatFormer is a tiny neural network that estimates heart rate from facial video by filtering motion noise in the frequency domain, and it can be trained without any labeled heart rate data. On cross-dataset motion tests it matches or beats much larger supervised models while using fewer than 15,000 parameters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's 0.1-BPM SOTA advantage is not controlled: single fixed seed, no error bars, and a CZT-based HR estimator that may not be applied to comparators; the claim needs a shared-protocol rerun.","rationale":"The reader's weakest-assumption analysis focuses on SCL's reliance on transformations preserving pulsatile content, which is a real concern for the unsupervised variant. However, the most load-bearing weakness for the central claim as stated is more immediate: the quantitative claim of beating RhythmFormer is built on a single-seed evaluation with no error bars and on an HR-estimation choice (CZT) that may not be shared by the comparators. The paper's own ablation shows CZT is worth several BPM, so a fair comparison must hold the HR estimator fixed. This is consistent with the reader's overall CONDITIONAL verdict and with their rationale mentioning missing code, error bars, and a clearly specified FFT baseline, but it shifts the primary locus from the SCL mechanism to the benchmark protocol. I therefore recommend keeping the verdict CONDITIONAL rather than moving it: the architecture and label-free loss are plausible and the ablations are informative, but the headline SOTA comparison and the SCL generalization assumption both require independent verification before endorsement.","tokens_in":15007,"tokens_out":10564,"duration_ms":122809,"concrete_test":"Re-run the cross-dataset benchmark with a single shared protocol for all methods: use the same preprocessing, the same 300-frame/10-frame-overlap evaluation, and compute HR from each method's predicted waveform with both FFT and CZT estimators. Include RhythmFormer (and ideally PhysMamba) from released checkpoints, and run BeatFormer-SL with at least 5 random seeds reporting mean ± std. If the PURE->MMPD MAE gap between BeatFormer-SL and RhythmFormer remains above 0.5 BPM under both HR estimators, the SOTA claim survives; if the gap shrinks below run-to-run noise or flips under FFT, the claimed advantage is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim, that BeatFormer-SL slightly beats RhythmFormer in cross-dataset MAE (8.85 vs 8.98 on PURE->MMPD, 8.98 vs 9.08 on UBFC-rPPG->MMPD), is not yet established because the comparison may not be apples-to-apples. Section 4.1 states that BeatFormer's predicted rPPG is filtered and HR is estimated with CZT, while the comparator numbers for RhythmFormer, PhysMamba, and others are taken from prior work that typically uses FFT-based HR extraction. This matters enormously: the paper's own ablation in Table 4 shows that replacing FFT with CZT changes MAE by roughly 4-5 BPM (14.50 to 9.15 without ZOCA; 13.07 to 8.85 with ZOCA), which is 30-40 times larger than the claimed 0.13 BPM advantage over RhythmFormer. If only BeatFormer benefits from CZT's spectral zoom, the headline gap could be an artifact of the HR estimator rather than of the learned filtering. Compounding this, the supervised model is trained for 20 epochs with a fixed random seed (§4.2), and no confidence intervals or multi-seed means are reported; a 0.1 BPM difference is plausibly within seed-to-seed and evaluation-pipeline noise. The SCL invariance concern raised by the reader is legitimate for the label-free variant, but it is downstream: the supervised head-to-head claim needs a controlled benchmark before the unsupervised mechanism can be meaningfully assessed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BeatFormer, a lightweight spectral transformer for remote photoplethysmography (rPPG) that operates on Chirp-Z Transform (CZT) spectra instead of raw temporal traces. The architecture combines a zoomed orthonormal complex attention (ZOCA) block with an energy-measurement feed-forward layer that filters RGB spectra. A second contribution, Spectral Contrastive Learning (SCL), uses hand-designed video transformations (HSV, LAB, temporal flipping, random occlusion) to train the model without PPG or heart-rate labels. The authors evaluate on PURE, UBFC-rPPG, and MMPD, reporting cross-dataset MAE of 8.85/8.98 BPM for the supervised model and 9.14/9.25 BPM for the label-free model, against 8.98/9.08 BPM for RhythmFormer, with about 14.86k parameters and 181.73 MACs. Ablations show that both CZT over FFT and ZOCA over plain complex attention improve MAE, and that the SCL loss approaches supervised performance.","tokens_in":15338,"tokens_out":4191,"duration_ms":49432,"significance":"If the reported results hold, the paper would demonstrate that a sub-15k-parameter spectral model can match state-of-the-art cross-dataset motion robustness and that a label-free training objective can come close to supervised performance. The design ideas are sensible: combining physiological priors from chrominance-based methods with a learned spectral filter is a plausible route to efficient and robust rPPG. The paper also benefits from ablations on the loss, the attention mechanism, the transform, and the SCL transformations, as well as from releasing code. The main limitations are that the headline comparison to prior work may not use the same heart-rate extraction pipeline, that no uncertainty quantification is provided, and that the label-free training assumption is not independently verified.","major_comments":[{"comment":"The central comparison may not be apples-to-apples. Section 4.1 states that BeatFormer's predicted rPPG is detrended, Butterworth-filtered, and converted to heart rate using CZT, but the comparator numbers in Table 1 are taken from prior publications that typically use FFT-based heart-rate extraction. This is load-bearing: Table 4 shows that replacing FFT with CZT changes MAE by roughly 4 to 5 BPM (14.50 to 9.15 without ZOCA; 13.07 to 8.85 with ZOCA), which is 30 to 40 times larger than the 0.13 BPM advantage claimed over RhythmFormer. The authors should rerun all baselines with the exact same evaluation pipeline, or at minimum report CZT-based MAE for the comparators and justify why their published numbers are directly comparable.","section":"§4.1, Table 1, Table 4"},{"comment":"The reported differences are not supported by uncertainty quantification. Section 4.2 specifies a fixed random seed, 20 epochs, and batch size 2, and no confidence intervals or multi-seed means are reported. The performance gaps at stake are 0.13–0.29 BPM, which is small relative to typical seed-to-seed variation in deep rPPG training, especially with small training sets like PURE and UBFC-rPPG. The authors should provide results over multiple seeds with means and standard deviations, and ideally significance tests over videos, before claiming that BeatFormer-SL outperforms RhythmFormer or that SCL is comparable to SL.","section":"§4.2, Table 1"},{"comment":"The SCL unsupervised claim rests on an unverified invariance assumption. The three assumptions in Section 3.3 state that the four video transformations preserve pulsatile frequency content while altering motion and other distortions, but this is asserted from physiological intuition rather than demonstrated. If the assumption fails in a new population, the label-free model has no corrective signal. The authors should provide empirical evidence, for example by showing that the predicted spectral peaks under each transformation agree with ground-truth heart rate on a labeled subset, or by measuring the consistency of the network's output across the four transformations. This is particularly important because Figure 4 shows that some transformations (e.g., spatial occlusion) behave differently across motion splits.","section":"§3.3, Figure 4"}],"minor_comments":[{"comment":"The statement that CZT has 'approximately 13 times higher frequency resolution than FFT' at 30 fps appears inconsistent with the stated zoom band of 0.66–2.5 Hz: the ratio of full-band width to zoom-band width is 30/(2.5-0.66) ≈ 16.3, not 13. Please clarify the calculation.","section":"§3.1"},{"comment":"The matrix form of CZT is notationally inconsistent: A is described as an N×N diagonal matrix while the displayed diagonal uses entries A^{-0}, A^{-1}, ..., A^{-n}, and W is called an M×N Vandermonde matrix while the displayed exponent pattern is unclear. Please align the indices and dimensions.","section":"§3.1, Eq. (3)"},{"comment":"The phrase 'with a fixed random seed' does not identify the seed value or whether the reported numbers are from that single seed. Please report the seed used and, preferably, multiple seeds with summary statistics.","section":"§4.2"},{"comment":"The computational-cost comparison says it follows the protocol of [82] and uses a FLOPs counting tool, but the table reports MACs. Please specify the input resolution and temporal length used for the count, and state whether the tool reports FLOPs or multiply-accumulate operations.","section":"§4.2.2, Table 2"},{"comment":"The 'Impact of video transformations in SCL training' paragraph refers to Figure 3, but the relevant panel appears to be Figure 4; please correct the cross-reference.","section":"§4.3"},{"comment":"The sentence 'denoting a notable performance improvement (approximately 3 BPM)' understates the effect shown in Table 4: the improvement is about 4.2 BPM with ZOCA and about 5.4 BPM without ZOCA. Please use the actual values.","section":"§4.3"},{"comment":"The text says the complex matrix multiplication develops 'four complex attention blocks', but Eq. (9) contains eight sign-combined terms. Please clarify how the four blocks correspond to the eight terms.","section":"§3.2.2, Eq. (6)–(9)"},{"comment":"There is a typo in the text: 'RythmFormer' should be 'RhythmFormer' to match Table 1 and the reference list.","section":"§4.2.1"}],"recommendation":"major_revision","confidential_remarks":"The main concern is the evaluation protocol rather than the method itself. If the authors can rerun comparators with their CZT-based heart-rate estimator and supply multi-seed statistics, the supervised claim would be credible; the SCL invariance issue could be addressed with an additional validation experiment. The paper fits the journal's scope and the code release is a positive factor."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — here's my take.\n\nBeatFormer is worth reading for one reason: it shows a 14.86k-parameter spectral rPPG model can nearly match supervised performance without any PPG or HR labels, using a contrastive objective (SCL) defined on physiologically motivated video transformations. That is a real, useful result, and the ablation study gives it credit: Table 4 shows CZT over FFT buys about 4–5 BPM of MAE, ZOCA over plain complex attention adds about 0.3 BPM, and SCL lands within about 0.3 BPM of supervised. Those directions are consistent with the design story.\n\nThe problem is the headline comparison to RhythmFormer. Section 4.1 states that BeatFormer's predicted rPPG is filtered and then HR is estimated with CZT. The comparator numbers are taken directly from prior papers, which standardly use FFT-based HR extraction. Given Table 4's own numbers, CZT vs. FFT changes MAE by 4–5 BPM—30 to 40 times larger than the claimed 0.1–0.13 BPM advantage over RhythmFormer. That means the advantage is confounded by the HR estimator. The paper needs a shared-protocol rerun where every method is evaluated with the same detrending, filtering, and HR estimation (ideally both FFT and CZT). Without that, \"slightly beats RhythmFormer\" is not supportable.\n\nSecond, there are no error bars or multi-seed means. Training uses 20 epochs and a fixed random seed. A 0.1 BPM difference is seed-level noise. Even after fixing the protocol, the authors need to report variance across seeds or splits.\n\nThird, the FFT baseline in Table 4 is underspecified. If FFT is evaluated over the full 30 Hz Nyquist band while CZT zooms into 1.84 Hz, the comparison conflates spectral resolution with the transform itself. A zero-padded FFT with matched bin spacing would isolate the zoom benefit.\n\nThe SCL invariance premise—that HSV/LAB conversions, temporal flip, and occlusion preserve pulsatile frequency content while disturbing motion—is plausible but untested. However, the cross-dataset MMPD result is independent external evidence that the learned representation tracks the true pulse, so I treat this as a secondary concern, not a dealbreaker.\n\nWho's this for: rPPG researchers interested in efficient frequency-domain models and label-free training. It deserves serious peer review, but as a major-revision candidate, not near-acceptance. A rerun with matched HR estimation, variance reporting, and a clear FFT baseline is the path.","headline":"Neat lightweight spectral rPPG architecture with honest ablations, but the SOTA claim is confounded by mismatched HR estimation and missing variance bars.","tokens_in":15877,"tokens_out":4982,"would_cite":true,"duration_ms":53230,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"BeatFormer claims a 14.86k-parameter spectral attention model can match state-of-the-art motion-robust remote heart rate estimation, and can be trained without any PPG or heart-rate labels.","keywords":["remote photoplethysmography","heart rate estimation","Chirp-Z Transform","spectral attention","complex transformer","contrastive learning","motion robustness","lightweight neural networks"],"falsifier":"Train BeatFormer-SCL on synthetic videos whose only periodic component is rigid head motion at a frequency inside the 0.66-2.5 Hz band, with no real cardiac signal; if the contrastive objective drives the network to output that motion frequency, the transformation-preservation assumption is false, whereas a flat or near-zero output would support it.","tokens_in":14800,"feed_emoji":"💓","tokens_out":5682,"duration_ms":57996,"temperature":0.7,"pith_summary":"The paper introduces BeatFormer, a remote heart rate estimation model built from a spectral filter transformer and the Chirp-Z Transform. Its supervised version reaches cross-dataset mean absolute heart rate errors of 8.85 BPM (PURE to MMPD) and 8.98 BPM (UBFC-rPPG to MMPD), slightly below the prior best model RhythmFormer while using 14.86 thousand parameters and 181.73 million MACs. The paper further shows a version trained with no PPG or heart rate labels, using a new spectral contrastive learning objective, stays close to the supervised result (9.14 and 9.25 BPM). If these numbers hold, they indicate that most of the parameter cost in rPPG networks is unnecessary and that explicit physiological priors in the frequency domain can replace labels. The motivation is that handcrafted chrominance methods are robust to motion but linear, while deep models are accurate but heavy and label-hungry; BeatFormer is proposed as the middle path.","feed_headline":"Label-free spectral model matches top remote heart-rate accuracy","feed_subtitle":"BeatFormer reaches 8.85 BPM cross-dataset error with 14.86k parameters, rivaling a 3.25M-parameter transformer.","key_machinery":"The central machinery is the Chirp-Z Transform (CZT) as a zoomable spectral front end, the Zoomed Orthonormal Complex Attention (ZOCA) block, and the Spectral Contrastive Learning (SCL) objective. The CZT evaluates the z-transform along a spiral contour so it can concentrate the same number of bins on the 0.66-2.5 Hz heart rate band that the FFT would spread over the full 30 Hz range, giving roughly 13x finer resolution on short windows. ZOCA is a complex multi-head attention whose weight rows are constrained (first row fixed to the luminance-cancelling [1,1,1]/√3 direction, remaining rows orthonormal) so that the model acts as a learned spectral filter with physiological priors instead of a free transformer. SCL is a triplet hinge loss on the squared Earth Mover's Distance between power spectra of transformed versions of the same video (positive) and different videos (negative), using HSV/LAB conversions, temporal flipping, and random occlusion as transformations that are assumed to preserve pulsatile frequency content while altering motion noise.","core_discovery":"BeatFormer is a lightweight transformer that operates directly on the frequency spectrum of spatially averaged facial skin RGB traces. It uses zoomed orthonormal complex attention (ZOCA), where the first row of each attention weight is fixed to the unit vector [1,1,1]/sqrt(3) and the remaining rows are constrained to be orthonormal, and an energy-measurement feed-forward that weights frequency subbands by their energy contribution relative to the input spectrum. The model is trained either supervised with a frequency-domain EMD loss plus orthonormality regularization, or without any labels using Spectral Contrastive Learning (SCL), which treats HSV/LAB color-space conversions, temporal flipping, and random occlusion of the same video as positive pairs and different videos as negatives. The paper's claim is that this combination reaches accuracy on the MMPD motion benchmark comparable to the best prior model while reducing the parameter count by over two orders of magnitude relative to RhythmFormer, and that SCL recovers most of the supervised gain without physiological labels.","pith_inferences":["A direct extension of this result is that the same label-free spectral contrastive objective could apply to other periodic physiological signals, such as respiration rate or pulse transit time variability, wherever a transformation set can be designed that preserves the target periodicity.","Because BeatFormer uses spatially averaged RGB traces rather than per-pixel features, its robustness presumably comes from spectral filtering of the average trace; this suggests pixel-level motion compensation and spectral filtering are partially interchangeable sources of motion robustness, a trade-off the paper does not directly measure.","The paper's claim that phase information helps disentangle pulse from motion could be tested by ablating the imaginary-part components of the CZT spectrum; if phase is truly load-bearing, an FFT-based magnitude-only variant should degrade substantially more than the roughly 3 BPM gap reported in the ablation table.","One implicit consequence is that the transformation list in SCL is not fixed by the method; a practitioner facing a new population or camera type could add transformations, such as color-temperature shifts or compression artifacts, and keep the training label-free as long as the cardiac frequency remains invariant."],"forward_implications":["If BeatFormer's cross-dataset numbers reproduce, rPPG models under 15k parameters can match the accuracy of transformer models with hundreds of times more parameters, shifting the cost profile of on-device heart rate monitoring.","The SCL result implies that a pulse signal can be learned from unlabeled facial videos, so training sets for rPPG need not come with synchronized PPG contact sensors, removing a major data-collection bottleneck.","The CZT's zoomed spectral resolution means short temporal windows of around 300 frames can resolve heart rate frequencies that FFT-based pipelines smear, which matters for real-time applications with low latency.","The orthonormality prior suggests the classic chrominance intuition, cancelling intensity via [1,1,1]/√3 and separating color directions, can be embedded as a structural constraint in a deep network rather than as a fixed handcrafted projection.","The paper's motion-split results indicate the gap between handcrafted and deep methods on walking, rotation, and talking scenarios can be closed by spectral rather than spatiotemporal feature learning."],"supporting_citations":[{"why":"Supplies the Chirp-Z Transform algorithm, the zoomable spectral analysis the whole model is built on.","marker":"[50]"},{"why":"Establishes adaptive spectral zoom in the heart rate band and sets the CZT size convention that BeatFormer adopts.","marker":"[11]"},{"why":"Provides the complex attention mechanism and min-max normalization that ZOCA is derived from.","marker":"[72]"},{"why":"Motivates the fixed first-row RGB weighting and the chrominance-based physiological prior for motion robustness.","marker":"[67]"},{"why":"Motivates energy-based filtering by showing pulsatile and motion signals have different relative RGB channel amplitudes.","marker":"[68]"},{"why":"RhythmFormer, the state-of-the-art baseline whose cross-dataset MAE BeatFormer matches or beats.","marker":"[82]"},{"why":"The MMPD benchmark dataset used for cross-dataset and motion-scenario evaluation.","marker":"[60]"},{"why":"Supplies the squared Earth Mover's Distance loss used in the SCL positive and negative pairing.","marker":"[18]"}],"fun_headline_variants":["No labels, 14k params, motion-robust heart rate: BeatFormer","Label-free spectral model 200x smaller, motion-robust heart rate","BeatFormer: tiny spectral, label-free, motion-robust HR","Cross-dataset heart-rate match without labels, 14k params"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the four video transformations, HSV and LAB conversions, temporal flipping, and random occlusion, always keep the true heart-rate frequency intact while changing everything the model should ignore, so that forcing spectral similarity within a video teaches the cardiac signal rather than some other stable artifact.","fun_headline_variants_meta":{"raw":{"variants":["No labels, 14k params, motion-robust heart rate: BeatFormer","Label-free spectral model 200x smaller, motion-robust heart rate","BeatFormer: tiny spectral, label-free, motion-robust HR","Cross-dataset heart-rate match without labels, 14k params"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001437,"raw_usage":{"total_tokens":5792,"prompt_tokens":945,"completion_tokens":4847,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":4765}},"tokens_in":561,"tokens_out":4847,"duration_ms":33603,"temperature":1.0,"reasoning_tokens":4765,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:45:33.772513+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train BeatFormer-SCL on synthetic videos whose only periodic component is rigid head motion at a frequency inside the 0.66-2.5 Hz band, with no real cardiac signal; if the contrastive objective drives the network to output that motion frequency, the transformation-preservation assumption is false, whereas a flat or near-zero output would support it.","supporting_citations":[{"cited_title":"The chirp z-transform algorithm","cited_arxiv_id":null,"evidence_quote":"Supplies the Chirp-Z Transform algorithm, the zoomable spectral analysis the whole model is built on."},{"cited_title":"Deep adaptative spectral zoom for improved remote heart rate es- timation","cited_arxiv_id":null,"evidence_quote":"Establishes adaptive spectral zoom in the heart rate band and sets the CZT size convention that BeatFormer adopts."},{"cited_title":"Complex transformer: A framework for modeling complex-valued sequence","cited_arxiv_id":null,"evidence_quote":"Provides the complex attention mechanism and min-max normalization that ZOCA is derived from."},{"cited_title":"Algorithmic principles of remote ppg","cited_arxiv_id":null,"evidence_quote":"Motivates the fixed first-row RGB weighting and the chrominance-based physiological prior for motion robustness."},{"cited_title":"Color-distortion filtering for remote photo- plethysmography","cited_arxiv_id":null,"evidence_quote":"Motivates energy-based filtering by showing pulsatile and motion signals have different relative RGB channel amplitudes."},{"cited_title":"RhythmFormer: Extracting Patterned rPPG Signals based on Periodic Sparse Attention","cited_arxiv_id":"2402.12788","evidence_quote":"RhythmFormer, the state-of-the-art baseline whose cross-dataset MAE BeatFormer matches or beats."},{"cited_title":"MMPD: Multi-Domain Mobile Video Physiology Dataset","cited_arxiv_id":"2302.03840","evidence_quote":"The MMPD benchmark dataset used for cross-dataset and motion-scenario evaluation."},{"cited_title":"Squared earth movers distance loss for training deep neural networks on ordered-classes","cited_arxiv_id":null,"evidence_quote":"Supplies the squared Earth Mover's Distance loss used in the SCL positive and negative pairing."}],"review_version":1}