{
  "pdf": "cross-view-psd-consistency-rppg.pdf",
  "title": "CROSS-VIEW PSD DISTILLATION FOR VIEWPOINT-ROBUST REMOTE PHOTOPLETHYSMOGRAPHY FARS Analemma",
  "elapsed": 53.7,
  "runs_mode": 1,
  "valid_runs": 1,
  "avg_score": 4.5,
  "scores": [
    4.5
  ],
  "score_std": 0,
  "final_verdict": "Reject",
  "final_confidence": 0.6,
  "conference_scores": {
    "soundness": 2.8,
    "presentation": 3,
    "contribution": 2.2,
    "overall_rating": 4.5,
    "confidence": 3
  },
  "strengths": [
    "Clear and well-motivated core insight: the PSD of the cardiac signal is view-invariant while visual appearance changes across viewpoints. This is a principled observation that justifies the frequency-domain distillation approach (Section 3.2, Eq. 1-2).",
    "Strong ablation demonstrating the necessity of asymmetric design: Table 2 shows that symmetric PSD (B) causes catastrophic frontal degradation (MAE 2.91→11.06 bpm, 2/3 seeds collapsed), while asymmetric with stop-gradient (B') achieves stable training and best performance. This is a convincing empirical argument for the design choice.",
    "Dual improvement on both frontal and side views: B' improves frontal MAE from 2.91→2.45 bpm and side average from 5.24→3.99 bpm, which is notable because distillation methods often sacrifice teacher performance for student gains (Table 1). The 34% gap reduction is a meaningful practical improvement.",
    "Statistical rigor in signal quality analysis: Section 4.6 reports a Wilcoxon signed-rank test (p=2.1×10⁻¹⁴, effect size r=0.56) and per-video improvement rate (66.7%), going beyond aggregate MAE comparisons."
  ],
  "weaknesses": [
    "Single-dataset evaluation on MCD-rPPG only, with side views limited to ~45° angles. The authors acknowledge this in the conclusion, but it severely limits generalizability claims. No evaluation on other multi-view datasets (e.g., M3PD mentioned in Related Work) or on more extreme viewpoints (60°–90°). The paper's title claims 'viewpoint-robust' rPPG but tests only one viewpoint offset.",
    "Baseline fairness concerns: The pooled + augmentation baseline (A+) had 2/3 training failures, and the symmetric PSD baseline (B) had 2/3 collapsed runs. These are not properly functioning baselines — they are strawmen. The paper lacks a strong multi-view training baseline that actually trains successfully (e.g., multi-view with separate batch normalization, progressive training, or established multi-task learning). The 'side-only' EfficientPhys (4.19 bpm side avg) is surprisingly competitive with B' (3.99 bpm), suggesting the improvement margin is modest.",
    "The contribution is incremental relative to existing knowledge: The stop-gradient mechanism is directly borrowed from BYOL/SimSiam (Chen & He, 2020) and Wang et al. (2022), and PSD-based losses for rPPG have been explored before (e.g., in frequency-domain supervision). The combination of these two ideas, while effective, does not constitute a significant methodological advance. The paper essentially applies a known technique (stop-gradient in Siamese learning) to a known domain insight (PSD is view-invariant).",
    "Limited evaluation metrics: Only HR MAE via FFT peak detection is reported. Standard rPPG evaluation includes RMSE, MAE for other metrics (e.g., SpO₂, respiration rate if available), and time-domain metrics like SNR or Pearson correlation of the predicted BVP waveform. Without these, it is unclear whether the improved MAE corresponds to genuinely better waveform quality or just better peak alignment."
  ],
  "must_fix_items": [
    "Add at least one more multi-view dataset for evaluation (M3PD is mentioned in Related Work and should be evaluated), or explicitly restrict claims to 45° side-view scenarios.",
    "Provide a properly functioning multi-view baseline — the current A+ baseline fails 2/3 runs, making the comparison uninformative. A simple multi-view training baseline (e.g., joint training with view-specific batch norm, or weighted loss balancing) that trains successfully is needed.",
    "Report additional standard rPPG metrics beyond HR MAE (e.g., RMSE, Pearson correlation, SNR of predicted waveform) to give a fuller picture of signal quality improvement."
  ],
  "runs": [
    {
      "run": 1,
      "score": 4.5,
      "verdict": "Reject",
      "confidence": 0.6,
      "strengths": [
        "Clear and well-motivated core insight: the PSD of the cardiac signal is view-invariant while visual appearance changes across viewpoints. This is a principled observation that justifies the frequency-domain distillation approach (Section 3.2, Eq. 1-2).",
        "Strong ablation demonstrating the necessity of asymmetric design: Table 2 shows that symmetric PSD (B) causes catastrophic frontal degradation (MAE 2.91→11.06 bpm, 2/3 seeds collapsed), while asymmetric with stop-gradient (B') achieves stable training and best performance. This is a convincing empirical argument for the design choice.",
        "Dual improvement on both frontal and side views: B' improves frontal MAE from 2.91→2.45 bpm and side average from 5.24→3.99 bpm, which is notable because distillation methods often sacrifice teacher performance for student gains (Table 1). The 34% gap reduction is a meaningful practical improvement.",
        "Statistical rigor in signal quality analysis: Section 4.6 reports a Wilcoxon signed-rank test (p=2.1×10⁻¹⁴, effect size r=0.56) and per-video improvement rate (66.7%), going beyond aggregate MAE comparisons."
      ],
      "weaknesses": [
        "Single-dataset evaluation on MCD-rPPG only, with side views limited to ~45° angles. The authors acknowledge this in the conclusion, but it severely limits generalizability claims. No evaluation on other multi-view datasets (e.g., M3PD mentioned in Related Work) or on more extreme viewpoints (60°–90°). The paper's title claims 'viewpoint-robust' rPPG but tests only one viewpoint offset.",
        "Baseline fairness concerns: The pooled + augmentation baseline (A+) had 2/3 training failures, and the symmetric PSD baseline (B) had 2/3 collapsed runs. These are not properly functioning baselines — they are strawmen. The paper lacks a strong multi-view training baseline that actually trains successfully (e.g., multi-view with separate batch normalization, progressive training, or established multi-task learning). The 'side-only' EfficientPhys (4.19 bpm side avg) is surprisingly competitive with B' (3.99 bpm), suggesting the improvement margin is modest.",
        "The contribution is incremental relative to existing knowledge: The stop-gradient mechanism is directly borrowed from BYOL/SimSiam (Chen & He, 2020) and Wang et al. (2022), and PSD-based losses for rPPG have been explored before (e.g., in frequency-domain supervision). The combination of these two ideas, while effective, does not constitute a significant methodological advance. The paper essentially applies a known technique (stop-gradient in Siamese learning) to a known domain insight (PSD is view-invariant).",
        "Limited evaluation metrics: Only HR MAE via FFT peak detection is reported. Standard rPPG evaluation includes RMSE, MAE for other metrics (e.g., SpO₂, respiration rate if available), and time-domain metrics like SNR or Pearson correlation of the predicted BVP waveform. Without these, it is unclear whether the improved MAE corresponds to genuinely better waveform quality or just better peak alignment."
      ],
      "must_fix_items": [
        "Add at least one more multi-view dataset for evaluation (M3PD is mentioned in Related Work and should be evaluated), or explicitly restrict claims to 45° side-view scenarios.",
        "Provide a properly functioning multi-view baseline — the current A+ baseline fails 2/3 runs, making the comparison uninformative. A simple multi-view training baseline (e.g., joint training with view-specific batch norm, or weighted loss balancing) that trains successfully is needed.",
        "Report additional standard rPPG metrics beyond HR MAE (e.g., RMSE, Pearson correlation, SNR of predicted waveform) to give a fuller picture of signal quality improvement."
      ],
      "conference_scores": {
        "soundness": 2.8,
        "presentation": 3,
        "contribution": 2.2,
        "overall_rating": 4.5,
        "confidence": 3
      }
    }
  ]
}