Lagrangian Turbulence Quality Guide & Auto-Research Roadmap for OpenPTV2¶

Executive Summary & Objective¶

Computer vision metrics such as Precision, Recall, Yield, and Frame-to-Frame Link Count quantify local association accuracy, but they do not measure real physical "quality" for Lagrangian turbulence research.

A tracking algorithm with $98\%$ frame-to-frame precision can still create $2\%$ false crossing swaps, generating unphysical velocity jumps $\Delta v \approx d_{\text{nn}} / \Delta t$ and massive acceleration spikes $a \propto 1/\Delta t^2$. These spurious spikes corrupt the non-Gaussian intermittency tails of the acceleration probability density function (PDF), invalidate velocity autocorrelation integrals $R_v(\tau)$, and ruin pair dispersion scaling $\langle \delta r^2(t) \rangle \sim g \epsilon t^3$.

This document defines the Lagrangian Turbulence Physics Quality Framework and outlines an Auto-Research Workflow leveraging ground-truth fluid trajectories (e.g., from the Johns Hopkins Turbulence Database - JHTDB) to evaluate and identify the ultimate tracker configurations balancing computational speed and real turbulent study quality.

1. Technical Computer Vision Metrics vs. Physical Turbulence Quality¶

Technical Metric	What It Measures	Why It Is Incomplete for Lagrangian Turbulence
Precision	$\frac{\text{TP}}{\text{TP} + \text{FP}}$	High precision does not guarantee physical derivative continuity. A single false link creates an extreme acceleration outlier that corrupts turbulent intermittency statistics.
Recall / Yield	$\frac{\text{TP}}{\text{TP} + \text{FN}}$	High recall with short, fragmented tracks ($L < \tau_\eta$, Kolmogorov scale) is useless for computing Lagrangian integral timescales $\tau_L$ or diffusion coefficients $D_L$.
Track Count	Total generated tracks	Splitting one long physical trajectory into 10 short fragments increases track count but destroys pair dispersion and velocity autocorrelation calculations.
Track Purity	Fraction of points from same true ID	Ignores gap-bridging capabilities across missing frames (e.g., laser sheet speckle or out-of-focus fade).

2. The 5 Core Physical Criteria for Lagrangian Turbulence Quality¶

To rigorously validate whether a tracker or hybrid cascading strategy produces physically sound trajectories, evaluation must incorporate the following five fluid mechanics criteria:

A. Trajectory Lifetime Distribution & Integral Scale Span ($\langle T \rangle / \tau_L$)¶

Physical Meaning: Lagrangian velocity autocorrelations $R_v(\tau) = \langle v(t)v(t+\tau)\rangle$ and structure functions $D_p(\tau) = \langle |v(t+\tau) - v(t)|^p\rangle$ require continuous trajectories spanning multiple Lagrangian integral timescales $\tau_L$.
Target Metric: Mean track duration $\langle T \rangle$, and the fraction of tracks exceeding $T > 10 \Delta t$, $T > 30 \Delta t$, and $T > \tau_L$.

B. Acceleration PDF Fidelity & Intermittency Kurtosis ($K_a$)¶

Physical Meaning: Fluid acceleration in intense turbulence is violently intermittent, characterized by heavy-tailed non-Gaussian PDFs with high kurtosis ($K_a = \langle a^4 \rangle / \langle a^2 \rangle^2 \approx 10 \dots 50$). Spurious track switches introduce artificial acceleration outliers that artificially inflate $K_a$.
Target Metric: Kurtosis Error Bias: $\Delta K_a = |K_{a,\text{pred}} - K_{a,\text{true}}|$.

C. Velocity Power Spectral Density (PSD) & Energy Cascade¶

Physical Meaning: In the inertial subrange, the Lagrangian velocity spectrum follows Kolmogorov scaling $E_L(\omega) \propto \omega^{-2}$.
Target Metric: High-Frequency Noise Floor. Spatial jitter and false links manifest as flat white noise at high frequencies $\omega > 1/\tau_\eta$.

D. Relative Pair Dispersion & Richardson Constant ($g$)¶

Physical Meaning: Pairs of fluid particles with initial separation $r_0$ undergo exponential separation (Batchelor regime), followed by cubic explosive dispersion $\langle |r(t) - r(0)|^2 \rangle = g \epsilon t^3$ (Richardson-Obukhov regime).
Target Metric: Pair Identity Swap Rate & Richardson Constant Error ($\Delta g$).

E. Gap-Bridging & Intensity Dip Resilience¶

Physical Meaning: Particle intensity drops below detection thresholds for $1 \dots 3$ frames due to laser sheet non-uniformity or out-of-focus motion.
Target Metric: Gap Re-link Recall Rate—the percentage of interrupted trajectories correctly re-identified after 1–3 missing frames without resetting particle index.

3. Auto-Research Validation Pipeline Architecture¶

To automate the discovery of optimal tracking strategies combining execution speed and turbulent physics quality, OpenPTV2 supports an end-to-end Auto-Research benchmark pipeline:

┌─────────────────────────────────────────────────────────────────────────┐
│              1. Ground Truth Generation (e.g. JHTDB)                    │
│   Direct Numerical Simulation (DNS) Direct Fluid Particle Trajectories  │
└────────────────────────────────────┬────────────────────────────────────┘
                                     │
                                     ▼
┌─────────────────────────────────────────────────────────────────────────┐
│              2. Synthetic Experiment Projection & Degradation           │
│   Add Gaussian Position Jitter, Intensity Dips, Out-of-Focus Missing    │
│   Detections, and Synthetic Ghost Particles                             │
└────────────────────────────────────┬────────────────────────────────────┘
                                     │
                                     ▼
┌─────────────────────────────────────────────────────────────────────────┐
│              3. OpenPTV2 Tracking Engine & Strategy Execution           │
│   • Single-Pass Engine (priority_segment_3d, nearest_hungarian_3d, etc.)│
│   • Hybrid Cascading Strategy 1 (Forward-Fast / Backward-Kalman)        │
│   • Hybrid Cascading Strategy 2 (Two-Scale Velocity Cascading)          │
└────────────────────────────────────┬────────────────────────────────────┘
                                     │
                                     ▼
┌─────────────────────────────────────────────────────────────────────────┐
│              4. Dual-Layer Performance & Physics Evaluator              │
│   • Layer A: Technical Metrics (Precision, Recall, Ghost%, ms/frame)    │
│   • Layer B: Lagrangian Physics Metrics (T/tau_L, ΔKa, PSD Noise Floor) │
└────────────────────────────────────┬────────────────────────────────────┘
                                     │
                                     ▼
┌─────────────────────────────────────────────────────────────────────────┐
│              5. Pareto Optimality Decision Engine                       │
│   Select Ultimate Strategy on Speed vs. Lagrangian Physics Pareto Front │
└─────────────────────────────────────────────────────────────────────────┘

4. Combined Quality Score Formulation¶

We define a holistic Lagrangian Turbulence Performance Score (LTPS):

$$\text{LTPS} = w_1 \cdot \text{Precision} + w_2 \cdot \text{PMT\%} + w_3 \cdot \min\left(1.0, \frac{\langle T \rangle}{20 \Delta t}\right) - w_4 \cdot \frac{|\Delta K_a|}{K_{a,\text{true}}} - w_5 \cdot \text{Ghost\%}$$

Where: - $w_1 = 0.25$ (Technical Precision) - $w_2 = 0.25$ (Perfect Match Trajectories) - $w_3 = 0.25$ (Normalized Mean Track Length) - $w_4 = 0.15$ (Acceleration Kurtosis Fidelity) - $w_5 = 0.10$ (Ghost Capture Penalty)

5. Roadmap for Future Auto-Research Campaigns¶

[ ] JHTDB Ingestion Interface: Connect openptv2.benchmarking to extract 3D Lagrangian trajectories from JHTDB Homogeneous Isotropic Turbulence ($Re_\lambda \approx 433$) or Forced MHD Turbulence datasets.
[ ] Synthetic Camera Projection: Project 3D DNS particles onto multi-camera 2D image planes using realistic optical calibration matrices and point spread functions (PSF).
[ ] Automated Parameter Sweeps: Run Bayesian optimization over tracker parameter spaces (dacc, gate_threshold, search_radius) using LTPS as the objective function.
[ ] Pareto Frontier Visualization: Generate publication-ready Pareto plots comparing Execution Time (ms/frame) vs. Lagrangian Physics Quality (LTPS).