Comprehensive Tracker Benchmarking, Selection & Strategy Guide for OpenPTV2¶

Executive Summary & Key Findings¶

Based on empirical cross-tracker benchmarking on turbulent synthetic flow fields, there is no single "best" tracker for every flow. For the current per-tracker recommendations see docs/trackers.md. (nearest_hungarian_3d and predictive_gmm_3d were removed on 2026-10-09, together with the hybrid cascading strategies built on them.)

1. Single-Pass Engine Benchmarks¶

Evaluated on turbulent flow dataset (Mean $N \approx 220$ particles/frame, $v = 1.52\text{ mm/frame}, d_{\text{nn}} = 6.98\text{ mm}, M = 0.218$):

Tracker Engine	Precision	Recall / Yield	Fragment Count ($F$)	Track Purity	Perfect Match %	Speed (ms/frame)
`priority_segment_3d` (Default)	0.974	0.901	3.56	0.970	73.7%	159.4 ms

2. Tracker Selection Decision Matrix¶

Condition-by-Condition Guide¶

Flow Condition	Recommended Engine / Hybrid	Primary Reason
High Density / Large Datasets ($N > 1,000$)	`priority_segment_3d`	$O(N)$ spatial grid cell indexing provides 4x–10x faster runtime without precision loss.

3. Parameter Tuning Reference¶

A. `priority_segment_3d` (Cython Engine)¶

dacc (Maximum Acceleration Window, mm):
Default: 5.5 mm
Tuning: Do not expand dacc arbitrarily! Expanding dacc from 5.5 to 50 drops precision from 0.974 to 0.810 due to enlarged candidate search bounds.
angle (Maximum Turn Angle, deg):
Default: 120.0 deg
Tuning: Decrease to 45.0–60.0 deg for directed laminar flows to prune improbable sharp turns.

4. Cheat Sheet & Execution Commands¶

Target Objective	Recommended Configuration / Strategy	Command
Default Fast Pass	Single Pass `priority_segment_3d`	`uv run python scripts/bench_trackers.py`
High Density / Large Scale	Spatial Grid Accelerated `priority_segment_3d`	`uv run python scripts/bench_trackers.py --density 5000`

5. Future Verification & Benchmarking TODO Backlog¶

To continuously validate and expand OpenPTV2 tracking strategies on experimental datasets, the following verification tasks are slated for future releases:

[ ] Adaptive Velocity Thresholding: Automatically infer dvxmax/dvxmin bounds per frame based on mean spatial displacement histograms prior to tracking.
[ ] Ensemble Consensus Voting: Implement an $N$-tracker consensus ensemble where a candidate link is accepted only if at least $M$ out of $N$ trackers agree.
[ ] GPU-Accelerated Bipartite Matching: Port global Hungarian cost matrix solvers to PyTorch/CuPy for $N > 50,000$ ultra-dense particle tracking.