Skip to content

Performance Evolution (Mar 2–16, 2026)

This document consolidates the superseded benchmark reports into a single optimization timeline, preserving the key numbers and design decisions that shaped the current architecture.

1. Initial Baseline (2026-03-02)

Dataset: ICRA 2020 forward, 50 images, 50 tags/frame. Hardware: Desktop CPU, Linux.

Stage Latency (ms) % Total
Preprocessing 4.53 8.1%
Segmentation 5.49 9.8%
Quad Extraction 6.22 11.1%
Decoding 39.49 70.6%
Pose Estimation ~0.21 <1%
Total 55.94

Key finding: Decoding consumed 70% of frame time (~0.8 ms/candidate). Homography estimation + bilinear sampling were the dominant costs.

2. SoA Migration (2026-03-03)

Migrated from discrete Quad objects to the DetectionBatch Structure-of-Arrays layout.

Stage Before (ms) After (ms) Speedup
Preprocessing 4.53 0.93 4.9x
Segmentation 5.49 1.80 3.1x
Quad Extraction 6.22 1.48 4.2x
Decoding 39.49 10.04 3.9x
Total 55.94 14.55 3.8x

Design decisions:

  • Eliminated L1 cache misses from object hopping via contiguous parallel arrays.
  • Enabled rayon parallelization on dense slices.
  • Per-tag decoding: 0.8 ms → 0.20 ms.
  • Partitioned pose estimation — heavy LM solver runs only on Valid markers.

Accuracy verification: Recall and RMSE matched pre-SoA baseline (Soft: 94.35% recall, 0.26 px RMSE).

3. Micro-Benchmark Baselines (2026-03-07 → 03-15)

Hardware: AMD EPYC-Milan, Linux, --release, single-threaded.

Two successive baselines captured the pre-SIMD state of per-stage latency:

Resolution Thresholding Segmentation Quad Extraction
VGA 1.2–1.3 ms 1.6–2.2 ms 3.8–4.8 ms
720p 3.3–4.1 ms 4.9–5.0 ms 12.7–16.7 ms
1080p 12.2–13.5 ms 9.2–11.7 ms 37.0–48.9 ms
4K 38.9–54.1 ms 29.7–38.9 ms 178–207 ms

Range reflects environmental variance across two measurement sessions; no code changes between them.

4. Hub Multi-Resolution Accuracy (2026-03-08)

Dataset: Hugging Face std41h12 (480p/720p/1080p). Hardware: AMD EPYC-Milan.

Validated 100% recall across resolutions. Key observations:

  • Reprojection RMSE grows with resolution (2.9 px at 480p → 4.8 px at 1080p) due to focal length scaling.
  • Rotation P50 improves at higher resolution (6.3° → 3.9°) — more pixels on edges.
  • Translation P50 excellent (~1 cm), but P99 tail grows at 1080p (~30 cm).

5. GWLF Introduction (2026-03-12/15)

Dataset: Hugging Face single_tag_locus_v1_tag36h11 (45–50 images/resolution).

Gradient-Weighted Line Fitting replaced ERF as the recommended corner refinement for robotics:

Metric ContourRdp+ERF GWLF Improvement
Det. RMSE (720p) 0.99 px 0.71 px 28%
Rotation P50 (720p) 2.22° 0.30° 7.4x
Precision 96.7% 100%
Latency (720p) 15.5 ms 13.8 ms 12% faster

Design decisions:

  • GWLF fits infinite lines to edge gradients via Weighted Orthogonal Distance Regression (PCA).
  • Corners computed as algebraic line intersections in homogeneous space.
  • Bilinear gradient sampling + Adaptive Transversal Windowing (\(\pm \max(2, 0.01L)\)).
  • Also established Fast vs Accurate pose mode comparison — both use Huber IRLS; Accurate adds gain-scheduled Tikhonov regularization.

6. DDA-SIMD Decoding (2026-03-16)

Replaced per-pixel matrix multiplications with a Digital Differential Analyzer + AVX2 gather-based bilinear interpolation:

Metric SoA Baseline DDA-SIMD Speedup
Batch (1024 tags) ~200 ms 0.063 ms ~3100x
Per-candidate ~200 µs ~62 ns ~3200x

Design decisions:

  • DDA eliminates O(n) matrix multiplications per pixel — uses incremental partial derivatives.
  • O(1) contrast gate (Fast-Path Funnel) rejects 70–80% of false-positive candidates.
  • has_simd_padding safety check prevents out-of-bounds 32-bit gathers.
  • Conclusion: Decoding is no longer the bottleneck — focus shifted to quad extraction and preprocessing.

7. Improvements Regression Verification (2026-03-15)

Verified that a large refactoring branch (misc/improvements) introduced no algorithmic regressions across all 15 regression tests (9 ICRA + 6 Hub). Changes included: error propagation (Result-based API), config validation, thread-local arenas, capacity hints, constant centralization, ~350 lines of dead code removal, and expanded benchmarks.

Per-tag pose latency: ~2.3 µs (Fast), ~4.7 µs (Accurate) at 50 tags.

Optimization Timeline Summary

Mar 02  55.94 ms/frame  ─── Initial baseline (decoding = 70%)
   │
Mar 03  14.55 ms/frame  ─── SoA migration (3.8x total)
   │
Mar 08  ── accuracy ──  ─── Hub multi-res: 100% recall validated
   │
Mar 12  ── accuracy ──  ─── GWLF: 7x rotation improvement
   │
Mar 16   0.063 ms/1024  ─── DDA-SIMD: decoding eliminated as bottleneck
   │
Mar 19  ── current ──   ─── SIMD CCL Fusion: 3x segmentation (see current docs)
   │
Mar 21  ── current ──   ─── EdLines GN + SOTA presets (see current docs)

Superseded Reports

The following individual reports were consolidated into this document:

  • baseline_20260302.md — Initial pipeline profile
  • soa_migration_20260303.md — SoA migration results
  • baseline_micro_20260307.md — First micro-benchmark baseline
  • baseline_micro_20260315.md — Second micro-benchmark baseline
  • hub_multi_res_20260308.md — Multi-resolution accuracy (std41h12)
  • hub_locus_v1_20260312.md — GWLF introduction + Fast/Accurate comparison
  • funnel_dda_20260316.md — DDA-SIMD decoding optimization
  • improvements_regression_20260315.md — Refactoring regression verification