Selecting Denoisers for Frozen Pedestrian Detectors Under Gaussian Noise and JPEG Acquisition: Degradation Matching over Restoration Fidelity

Technical Summary: Selecting Denoisers for Frozen Pedestrian Detectors Under Gaussian Noise and JPEG Acquisition

Problem Statement
In intelligent transportation systems (ITS), pedestrian detectors are often deployed as "frozen" components—pre-trained, certified, and fixed weights that cannot be retrained for specific field conditions. When these detectors encounter sensor noise (e.g., additive white Gaussian noise) and acquisition artifacts (e.g., JPEG compression) during inference, their performance degrades sharply. The standard defense is to place a denoiser before the frozen detector. However, practitioners typically select denoisers based on restoration fidelity metrics (PSNR, SSIM) or assume that "blind" all-in-one models (trained on diverse corruptions) will universally outperform specialized ones.

This paper challenges the assumption that higher restoration fidelity implies better downstream detection performance for frozen pipelines. It posits that for a fixed detector and a fixed noise level, the critical variable is degradation-model match: whether the denoiser's training corruption profile aligns with the specific degradation pipeline (Gaussian noise + JPEG re-encoding) present in the deployment.

Methodology
The study employs a controlled experimental framework using a frozen YOLOv9-M pedestrian detector and a fixed noise grid ($\sigma \in \{10, 20, 30\}$) with an optional JPEG re-encoding stage ($q=95$).

1. Denoiser Roster: The authors evaluate a two-tier roster of denoisers:
   • Matched: Models trained specifically on the pipeline's corruption (e.g., $\sigma$-conditioned SwinIR, FFDNet, BM3D, and own-trained DnCNN/CAE).
   • Mismatched (Blind/All-in-One): Models trained on broad corruption sets or real-world noise but not the specific Gaussian+JPEG recipe (e.g., Restormer, PromptIR, NAFNet).
2. Structural Preservation Score (SPS): To move beyond pixel fidelity, the authors introduce a gradient-based metric, SPS, which measures the Pearson correlation between the Sobel gradient maps of the denoised image and the clean reference. This metric operationalizes the Structural Preservation Hypothesis (SPH): a denoiser is valuable if it preserves the edges and gradients the detector relies on, rather than maximizing pixel-level accuracy.
3. Calibration and Residual Flagging:
   • A within-$\sigma$ calibration relation is fitted on matched denoisers linking SPS to detection gain ($\Delta d$).
   • Mismatch Flag: Mismatched (blind) denoisers are expected to fall below this calibration relation. The standardized residual ($z$) of a candidate's performance from the fitted line serves as a flag: a negative $z$ indicates the denoiser is under-performing relative to its structural score, signaling a degradation mismatch.
4. Controlled Ablation: To isolate the cause of performance drops, the JPEG stage is toggled on and off. This tests whether the failure of blind denoisers is due to the specific mismatch with the JPEG artifact rather than general incapacity.
5. Routing Algorithm: An offline-calibrated routing table (Algorithm 1) uses a wavelet-based noise estimator ($\hat{\sigma}$) to select the cheapest eligible denoiser (lowest latency) that passes the residual flag, requiring no detector runs at deployment time.

Key Results
The study presents evidence across four datasets (PnPLO, CrowdHuman, CityPersons, INRIA) and nine frozen detectors:

• Degradation Match vs. Fidelity: Under a Gaussian-plus-JPEG recipe, $\sigma$-matched denoisers recover detection significantly (e.g., at $\sigma=30$ on PnPLO, matched SwinIR achieves $\Delta d \approx +0.238$, while blind Restormer/PromptIR achieve $\approx +0.011$). Conversely, blind denoisers often fail to recover detection despite having PSNR scores comparable to the weakest matched baselines.
• The Failure of Fidelity Metrics: PSNR and SSIM correlate well with detection gain for matched denoisers but mislead for mismatched ones. Blind denoisers often sit above the fidelity calibration relation (positive residuals), falsely suggesting they are beneficial, while their actual detection gain is near zero.
• SPS as a Discriminator: While PSNR has a higher average correlation with $\Delta d$ on the matched set, the SPS residual successfully flags mismatched denoisers. At $\sigma=30$, the SPS residual for blind models is negative (exposing the shortfall), whereas PSNR residuals are positive (concealing the failure).
• Ablation Evidence: Removing the mismatched JPEG stage from the pipeline restores the performance of frozen blind denoisers (Restormer $\Delta d$ jumps from $+0.000$ to $+0.380$), confirming that the degradation mismatch, not the model architecture alone, drives the failure.
• Generalization: The separation between matched and blind denoisers reproduces ordinally on CrowdHuman and CityPersons (driving scenes) under the same JPEG recipe. The effect vanishes on pure-Gaussian INRIA data (no JPEG) and reappears immediately when JPEG is added, proving the boundary is the acquisition recipe, not the dataset.
• Cost Efficiency: A lightweight, $\sigma$-matched denoiser (FFDNet, 17.5 ms) achieves recovery comparable to heavy transformer models (SwinIR, 3130 ms) at roughly 180x lower computational cost.

Significance and Claims
The paper claims that for frozen pedestrian detectors in ITS, degradation-model match is a stronger determinant of detection utility than restoration fidelity.

• Rejection of "Better Restoration = Better Detection": The study demonstrates that selecting denoisers based on PSNR/SSIM or blind all-in-one models can misdirect selection, either concealing non-recovery or flagging harm where none exists.
• Operational Rule: The authors propose a deployable recipe: estimate the noise level, route frames through an offline-calibrated table, and use the SPS residual sign to screen out mismatched restorers. This allows for the selection of lightweight, cost-effective denoisers without retraining the detector.
• Modesty of Scope: The authors explicitly state that the effect is regime-contingent (dependent on the presence of JPEG compression) and that the amplitude of recovery depends on the "headroom" available in the specific dataset (how much the benchmark collapses under noise). They do not claim universality across all degradation types (e.g., motion blur, haze) or detector architectures (e.g., DETR), noting these as future work. The findings confirm and quantify the qualitative observation by Sun et al. [6] for the specific case of frozen pipelines.