Vision intelligence, optimized.

Adaptive computer-vision experiments with evidence-backed decisions.

Oct 25, 2026 | Session #01
Edge Detection Tuning
BEST QUALITY
100.0%
▲ +100% Edge F1 Score (↑ better)
LATENCY
1.18 ms
▼ -20.3% Median per image (↓ better)
EXPERIMENTS
05
3 valid / 2 rejected
IMPROVEMENT
+25.5%
Composite Score

Live Experiment

Agent is iteratively tuning OpenCV 5 parameters based on visual feedback and metrics.

Iteration 05 / 05 Ready
BASELINE Run default parameters 1.48 ms | 0.00 F1
EVALUATE Analyze output quality 1.00 edge recall
ADJUST PARAMETERS Modify Canny thresholds low: 25, high: 50
VALIDATE Run & compare metrics Checking...
⚡ Optimization Ready. Select a preset or upload an image to begin.

Agent reasoning

Edge recall below target (0.00 < 0.70) Detected zero edges; subtle low-contrast boundaries missed
14:28:31
Lowering Canny thresholds aggressively low: 50 → 20, high: 150 → 52 (recover subtle boundaries)
14:28:32
Testing next configuration Running OpenCV 5 Canny with updated parameters
14:28:33
Evaluating output... Computing quality metrics (F1, edge density, noise)
14:28:33

Visual comparison

Input image, baseline output, best validated result, and reference (ground truth).

Image 4/5
INPUT IMAGE
Input Original Image
BASELINE EDGES
Baseline Output
BEST VALIDATED
Selected
Best Validated Candidate
REFERENCE (GROUND TRUTH)
Ground Truth Edge Map

Quality vs. Latency

Trade-off between output quality and inference time across experiments.

Valid Rejected Baseline
1.0 0.8 0.6 0.4 0.2 1 10 100 1000 Latency (ms, log scale) Quality (Edge F1)

Experiment history

All experiments with parameters, metrics, and decisions.

# PREVIEW PARAMETERS (CANNY) LATENCY EDGE F1 Δ VS BASELINE STATUS

Experiment Archive & Benchmark Runs

Complete history of all autonomous optimization loops and AWS cloud traces.

Curated Benchmark Datasets

Standardized static image test suite with verified ground-truth reference edge maps.

Grounded Evaluation Methodology

Mathematical formulation of Ground Truth F1, morphological tolerance dilation, and proxy metrics.

1. Ground Truth F1 Score

For images with verified ground-truth references, we calculate Precision and Recall using a 1-pixel morphological tolerance window (BSDS standard practice):

Precision = |E_cand ∩ Dilate(E_ref, 1px)| / |E_cand| Recall = |E_ref ∩ Dilate(E_cand, 1px)| / |E_ref| F1 = 2 × (Precision × Recall) / (Precision + Recall)

2. Unlabelled Proxy Quality

For unlabelled images where reference edge maps are unavailable, three proxy metrics prevent degraded configurations:

  • Edge Continuity: Connected components ratio penalizing fragmented edge dust.
  • Noise Ratio: Fraction of isolated single-pixel edge noise.
  • Density Fitness: Checks that edge density stays within natural boundaries (0.5% - 10%).

3. High-Precision Timing Harness

To eliminate CPU thermal throttling and operating system scheduler jitter:

T_candidate = Median(Pass_1, Pass_2, ..., Pass_5) after 3 Warmup passes

Decision Audit Trail

Chronological, immutable audit record of perception-action cycles and parameter transitions.

Loading audit trail...

System & Cloud Configuration

OpenCV 5 engine status, AWS compute credentials, and deployment telemetry.

Direct HTTPS POST endpoint for remote cloud execution