Dctracks Track Recon Eval

dctracks-track-recon-eval

DCTracks: An Open Dataset for Machine Learning-Based Drift Chamber Track Reconstruction — Qian Liyan et al. (2026) (arXiv:2602.14571, 2026)

What this evaluates

Evaluates the performance of machine learning and traditional algorithms for reconstructing particle tracks in drift chamber detectors. It probes hit-level matching accuracy, track-level reconstruction efficiency, charge identification correctness, and momentum resolution under realistic detector conditions.

Datasets

• DCTracks — total ?; splits: test (-1); repo https://github.com/lyqian1220/DCTracksMetrics.git

Metrics

• hit efficiency — range: [0, 1]
  • Fraction of a particle's detectable truth hits that are correctly reconstructed and matched to that particle: epsilon_hit = N_hit_matched / N_hit_detectable.
• hit purity — range: [0, 1]
  • Fraction of reconstructed hits assigned to a track that are correctly matched to the originating particle: p_hit = N_hit_matched / N_hit_assigned.
• track efficiency (primary) — range: [0, 1]
  • Fraction of detectable truth tracks (>=6 detectable hits) for which a matched reconstructed track exists: epsilon_track = N_track_matched / N_track_detectable.
• track charge efficiency — range: [0, 1]
  • Fraction of detectable truth tracks reconstructed with the correct charge: epsilon_track,q = N_track_matched,q-correct / N_track_detectable.
• wrong charge rate — range: [0, 1]
  • Fraction of detectable truth tracks reconstructed with the wrong charge: R_wrong,q = N_track_matched,q-incorrect / N_track_detectable.
• clone rate — range: [0, 1]
  • Total count of clone tracks (multiple reconstructions matching the same truth track) divided by total detectable truth tracks: R_clone = N_track_clone / N_track_detectable.
• fake rate — range: [0, 1]
  • Total number of fake tracks (fails purity/efficiency criteria) divided by total detectable truth tracks: R_fake = N_track_fake / N_track_detectable.
• pT resolution — range: percent
  • 68% coverage of the absolute normalized residual distribution around its median: r(pT) = P_68%(|eta_pT - P_50%(eta_pT)|), where eta_pT = (pT_reco - pT_MC) / pT_MC.

Input / output format

Input: Detector readout hits (after digitization, thresholding, and inefficiency losses) and Monte Carlo truth information for single- and two-track events.

Output: Reconstructed tracks with assigned hits, charge, and kinematic parameters (e.g., transverse momentum pT).

Scoring recipe

# 1. Identify detectable truth tracks (>=6 detectable hits)
# 2. For each reconstructed track, compute hit purity and hit efficiency
# 3. A track is "matched" if purity>0.5, hit_eff>0.2, and matched_hits>=6
# 4. If multiple match same truth track, keep highest hit_eff (others are clones)
# 5. Tracks failing criteria are "fake"
# 6. Compute rates:
track_eff = len(matched_truth_tracks) / len(detectable_truth_tracks)
hit_eff = sum(matched_hits_per_truth) / sum(detectable_hits_per_truth)
hit_purity = sum(matched_hits_per_truth) / sum(all_assigned_hits)
# 7. Charge metrics: fraction of matched tracks with correct/incorrect charge
# 8. pT resolution:
residuals = [(r.pt - t.pt) / t.pt for r, t in matched_pairs]
pT_res = quantile(abs(residuals - median(abs(residuals))), 0.68)

Common pitfalls

• Confusing hit-level metrics with track-level metrics; the paper strictly separates them and defines matching criteria at the track level.
• Misinterpreting 'clone tracks' as fake tracks; clones are valid matches that share a truth track, while fakes fail purity/efficiency thresholds.
• Using standard deviation for pT resolution instead of the 68% coverage around the median, as specified for potentially non-Gaussian distributions.

Evidence (verbatim from paper)

Track efficiency ($\epsilon_{\text{track}}$) is defined as the fraction of detectable truth tracks for which a matched reconstructed track exists: $\epsilon_{\text{track}}=\frac{N_{\text{track}}^{\text{matched}}}{N_{\text{track}}^{\text{detectable}}}$. Here, we define a simulated particle a detectable truth track if it has at least six detectable truth hits. $N_{\mathrm{track}}^{\mathrm{detectable}}$ denotes the number of detectable truth tracks in the samples and $N_{\mathrm{track}}^{\mathrm{matched}}$ denotes the subset that have a matched reconstructed track.

Citation

@misc{qian2026dctracks,
  title={DCTracks: An Open Dataset for Machine Learning-Based Drift Chamber Track Reconstruction},
  author={Qian Liyan et al. (2026)},
  year={2026},
  note={arXiv:2602.14571}
}

• arXiv: 2602.14571