Locus Release Performance Report (2026-03-22)
This report documents the performance of Locus v0.2.6 prior to release, 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.
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% | 0.2870 | 44.95 |
| Locus (Hard) | 76.87% | 0.2567 | 33.97 |
Synthetic Throughput
Measured on procedurally generated images (1080p).
| Tag Count | Latency (ms) | Recall |
|---|---|---|
| 1 Tag | 8.75 | 100.0% |
| 10 Tags | 5.45 | 100.0% |
| 50 Tags | 8.68 | 100.0% |
| 100 Tags | 11.76 | 100.0% |
Tier 2: Micro-Benchmarking (Divan)
Fine-grained mathematical kernels measured strictly single-threaded on 1080p images.
Mathematical Kernels
| Kernel | Median Latency |
|---|---|
| Thresholding (Apply) | 2.74 ms |
| Integral Image | 1.99 ms |
| Segmentation (1080p) | 2.47 ms |
Decoding & Pose
| Operation | Latency |
|---|---|
| Dictionary Lookup (36h11) | 372.8 ns |
| Soft Decoding (200 cands) | 161.7 µs |
| Pose Estimation (50 tags) | 77.8 µs |
Tier 3: Rust Regression Suite
ICRA 2020 (SOTA Metrology)
| Dataset | Recall | RMSE (px) | Latency (ms) |
|---|---|---|---|
| Forward (Pure Tags) | 96.23% | 0.3152 | 41.53 |
| Forward (Checkerboard) | 91.43% | 0.4577 | 57.89 |
| Circle (Pure Tags) | 83.27% | 0.3931 | 56.53 |
| Circle (Checkerboard) | 76.01% | 0.4171 | 56.18 |
Hub Regression (Rendered Tags)
Measured on the Hugging Face Hub benchmarking datasets.
| Dataset (36h11) | Mode | Recall | RMSE (px) | Latency (ms) |
|---|---|---|---|---|
| 640x480 | SOTA | 100.0% | 0.2112 | 10.12 |
| 720p | SOTA | 100.0% | 0.2456 | 21.45 |
| 1080p | SOTA | 100.0% | 0.2767 | 45.32 |
| 4K (2160p) | SOTA | 100.0% | 0.3124 | 168.4 |
| 1080p | Fast | 100.0% | 0.3542 | 12.15 |
Technical Updates in this Version
- Unified Config: Python CLI now correctly respects all Rust
DetectorConfigfields, includingquad_max_elongationandquad_min_density. - Soft Decoding Parity: Verified that Python
DecodeMode.Softmatches Rust SOTA results (96.23% recall on ICRA forward). - Benchmarking Consistency: Updated
LocusWrapperto pass all high-fidelity parameters to the underlying engine.