Boundary search
Find the minimum parameter value that destabilizes a particular object.
Know the frame
before it fails.
CVFuzz exposes the precise condition where your vision model loses its footing—then gives you the evidence to recreate it.
EXPOSURELOW LIGHTMOTION BLURDEFOCUSCOMPRESSIONFOGOCCLUSIONGLARE
Accuracy is a summary. Reality is a sequence.
The edge is where
the story starts.
Aggregate scores can make a model feel invincible. But a shift in weather, light, or motion can erase a detection in a single frame.
CVFuzz searches that moment deliberately. It changes one condition at a time, finds the smallest breaking change, and preserves everything needed to see it again.
One clean path
to the boundary.
- I
Load
Start with the scene.
Bring your YOLO model, a video, and a versioned YAML configuration. Nothing leaves your machine.
- II
Probe
Change one variable.
CVFuzz renders realistic conditions and searches the severity where a detection becomes unstable.
- III
Prove
Keep the moment.
Review the frame, threshold, output, and configuration. The result is evidence, not a mystery.
Watch confidence
come apart.
Compare original and transformed streams in lockstep. Follow confidence, target retention, class changes, and localization drift down to the exact frame.
- confidence loss
- −57%
- first boundary
- 11px
- target retention
- 43%
Built to make
instability legible.
Stream evaluation
Evaluate original and transformed video as synchronized, inspectable stages.
Clear signals
Measure missed objects, confidence collapse, class changes, and localization drift.
Local evidence
Keep manifests, videos, metrics, and settings together—without a database.