Skip to content

Locus Release Performance Report (2026-04-18)

This report documents the performance of Locus v0.3.1, covering end-to-end latency, mathematical kernel efficiency, and regression accuracy.

Tier 1: End-to-End Performance (Python CLI)

Validated on the ICRA 2020 dataset and synthetic targets. Measurements taken with RAYON_NUM_THREADS=1.

Real-World (ICRA 2020 - Scenario: forward)

Measured on all 50 images of the forward/pure_tags_images subset.

Configuration Recall RMSE (px) Latency (ms)
Locus (Soft) 96.23% 0.3152 151.12
Locus (Hard) 76.87% 0.2567 92.24

Note: Python CLI latency includes significant FFI and result-marshalling overhead. Rust core latencies for the same workload are ~32ms (Hard) and ~42ms (Soft).


Tier 2: Micro-Benchmarking (Divan)

Fine-grained mathematical kernels measured strictly single-threaded on 1080p images. Substantial improvements observed compared to the March 2026 baseline.

Mathematical Kernels (1080p)

Kernel Latest Median (ms) Baseline (2026-03-19) Speedup
Thresholding (Apply) 1.51 ms 4.26 ms 2.82x
Segmentation 2.57 ms 2.95 ms 1.15x
Quad Extraction 23.79 ms 34.35 ms 1.44x

Tier 3: Rust Regression Suite

ICRA 2020 (HighAccuracy Presets)

Scenario Preset Recall RMSE (px) Latency (ms)
Forward Standard (Soft) 96.23% 0.3152 41.81
Forward Grid 91.43% 0.4577 58.53
Forward HighAccuracy 46.31% 0.7535 15.43

Investigation: HighAccuracy Performance

HighAccuracy exhibits lower recall on ICRA 2020 compared to Hub datasets. - ICRA 2020 (Forward): 46.3% Recall - Hub (1080p): 95.6% Recall

Root Cause: Sub-Pixel Bit Density. Analysis of the ICRA 2020 dataset reveals that a significant portion of tags are extremely small relative to their bit grid: - Forward: ~18% of tags have < 1.2 pixels per bit (PPB). - Circle: Thousands of tags have < 1.0 PPB (some as low as 0.03 PPB).

The HighAccuracy preset is optimized for maximum precision on high-quality images. It uses EdLines quad extraction and disables Laplacian sharpening to avoid distorting the Point Spread Function (PSF). - EdLines sensitivity: Empirical testing shows EdLines recall collapses on tags with < 1.5 PPB (reaching < 1% for < 1.2 PPB), whereas the legacy ContourRdp algorithm remains robust down to ~1.2 PPB. - Sharpening impact: Disabling sharpening (as in HighAccuracy mode) further degrades recall on small tags by smearing bit boundaries. Enabling sharpening recovers EdLines recall to 100% for tags > 2.0 PPB, but it remains ineffective for the "sub-pixel" tags prevalent in ICRA 2020.

For real-world robotics tracking where high recall on distant tags is required, the standard profile (Soft Decode + ContourRdp + Sharpening) remains the recommended configuration. The high_accuracy profile should be reserved for high-resolution near-field calibration where PPB is > 5.0.

Hub Regression (Rendered Tags - 1080p Tag36h11)

Mode Recall Trans P50 (mm) Latency (ms)
Standard 100.0% 5.3 mm 21.69
HighAccuracy 95.56% 0.7 mm 11.44

Technical Updates in this Version

  • Snapshot Stabilization: All regression snapshots have been updated to established v0.3.1 baselines.
  • Kernel Optimization: SIMD-accelerated thresholding and segmentation provide significant throughput gains.
  • Metrology Handover: GN corner covariances now propagate directly to the weighted LM pose solver in HighAccuracy mode.