Proprietary recognition engine

Built to understand
your specifics.

Your objects. Your distinctions. Your operating conditions. Vektra brings them together in a recognition system designed around the work you need to do.

One engine. Purpose-built layers.

Beyond a generic label.

The value is in the distinctions: one component from another, an expected configuration from a condition that needs review, an asset class from its application-specific variants.

Vektra brings custom recognition models, domain context and output logic into a coordinated software architecture.

Object recognitionVisual classificationCustom taxonomiesStructured evidence
01 / Inputs

An observation enters.

Video, imagery and streams arrive hashed, with their timestamps intact, so everything that follows can be traced to a frame.

02 / Detect

Independent detectors look.

Detectors run alone or in tandem on the same frame. Their raw observations are kept apart; agreement is a signal to inspect, not a verdict.

03 / Verify

A second look at original resolution.

Each candidate is cropped from the original pixels and re-examined by a separately trained verification layer. Its score stays separate.

04 / Associate

Persistence across frames.

Same-class detections are associated across frames, so a vehicle faint in one frame is confirmed over several and a single frame never decides.

05 / Layers

Your vocabulary, your logic.

Domain layers map the observation onto the classes, attributes and distinctions you defined, and application logic decides what it means for the next step.

06 / Outputs

A record a reviewer can defend.

Asset class, context and evidence are written as structured records with the model, the frame and the person behind each one.

Inside the pipeline

Seven stages.
Each one accountable.

Every stage writes its own evidence. A detection, a verification, a visual match and a human decision are separate records, so any answer can be traced to the frame, the model and the person behind it.

See the capabilities
Engineered for accuracy

Make missed detections
part of the design brief.

Accuracy depends on the scene, the classes and the consequences of a miss. We build the evaluation around all three.

Domain-specific data

Representative examples and challenging cases shape the model around actual environmental variation, object appearance and scene complexity.

Class-level evaluation

Assess recognition by class and condition, with explicit attention to false negatives, false positives and ambiguous observations.

Controlled improvement

Review difficult examples, refine model layers and release updates against defined acceptance criteria and versioned evaluations.

Model engineering

How a model layer
is built.

The discipline behind every Vektra model layer, from the first annotated frame to a released engine.

Viewpoint-specific data

Most of the training set comes from the cameras, angles and conditions the model will actually face: deliberately biased toward the deployment rather than the internet.

Annotation discipline

Fixed label sets, tight boxes, a minimum-visibility rule for partly hidden objects, and human review of every label.

Hard negatives

Look-alike objects and empty scenes are collected on purpose, because false alarms come from whatever resembles the target.

Frozen validation

A golden validation set never enters training. Splits are grouped by recording and camera so nothing leaks between training and test.

Class-level metrics

Per-class precision and recall, false alerts per camera-hour, short-event recall and latency percentiles, compared on the same held-out set for every release.

Gated release

Training runs need explicit authorization, starting weights are hash-pinned, and no model is promoted automatically.

Beyond the public baseline

Where the field is.
Where Vektra goes further.

Open detectors are a starting point, not a product. Vektra engineers the layers that public models leave to the integrator.

CapabilityTypical open detectorVektra engine
ArchitectureSingle-frame, grid-based detection tuned to internet benchmarksGlobal-context detection, multi-scale small-object layers and multi-frame tracking tuned to your sensors
TaxonomyFlat generic labels: car, truck, airplaneHierarchical classes down to type, variant, configuration and markings, defined with you
SensorsVisible imagery onlyVisible, thermal and multi-band inputs with cross-band agreement
UncertaintyA raw score and a thresholdCalibrated confidence, an explicit unknown class and a decoy and look-alike assessment
EvaluationMean precision on a public datasetPer-class, per-condition evaluation on your environment, with missed detections tracked as the primary risk
VerificationOne pass, one scoreIndependent detectors, a separate verification layer at original resolution and geometric proof for visual matches
IntegrityWeights loaded from wherever they were downloadedHash-attested models, immutable observations and an append-only audit trail
OperationCloud inference, dependent on a linkEdge, server or hybrid, with full recognition held locally when links are denied
Where generic recognition fails

We design for the cases
that break ordinary detectors.

Most recognition failures are predictable. They cluster around scale, shift, clutter, concealment, sensor physics and the link to the operator. We treat each as a design requirement.

Read the field note
01 / SMALL OBJECTS

Targets that are a few pixels wide.

Wide-area drone and satellite frames put most objects of interest under 32 pixels. Generic detectors trained on close-range imagery lose them in the noise. Vektra builds multi-scale model layers and tracks across frames, so a vehicle that is faint in one frame is confirmed over several.

Four small vehicles at a track junction as a detector working at a reduced input scale receives them: a few blurred blocks Detector input scale
The same crop at native resolution: a pickup, two trucks and a tractor parked at the junction Native pixels
02 / DOMAIN SHIFT

Trained in one world. Deployed in another.

Published benchmarks show detectors losing a large share of their accuracy when climate, terrain, season or sensor changes, with missed detections growing fastest. Vektra trains and evaluates on your conditions, then re-evaluates each release against them.

A logistics yard in clear late-afternoon light: white trailers, containers and a service van, the conditions a model was trained in Training conditions
A logistics yard under heavy snow with falling snow and no shadows, the conditions it is deployed in Deployment conditions
03 / CLUTTER FALSE POSITIVES

A shadow that reads as a vehicle.

Grid-based detectors reason locally, so shadows, rubble and scrap routinely trigger false alarms. Vektra combines global-context architectures with calibrated confidence and an explicit unknown class, so the system says what it does not know.

A scrap yard at a detector’s input scale: debris piles, long shadows and a flatbed truck reduced to similar dark blocks Detector input scale
The same crop at native resolution: a flatbed truck beside vehicle-shaped debris piles with long shadows Native pixels
04 / CAMOUFLAGE & CONCEALMENT

Nets, treelines, dazzle and decoys.

Concealment is designed to defeat both eyes and models. Vektra fuses visible and thermal inputs where available, reasons over shape, context and motion rather than texture alone, and treats decoy discrimination as a first-class evaluation.

Vehicles under camouflage netting at a treeline, from directly above, at a detector’s input scale Detector input scale
The same crop at native resolution: three netted utility vehicles and a trailer under the trees Native pixels
05 / THERMAL CONTRAST

Hot engine. Cold engine. Same target.

Infrared is strong against foliage and light fog, and weak when a target matches its background. Vektra models learn per-band signatures and use cross-band agreement, so a loss of contrast in one band does not become a loss of the target.

Thermal frame of a vehicle park in the white-hot palette: one car with a warm engine beside cold vehicles that match the ground White hot
The same thermal crop in the black-hot palette Black hot
06 / DENIED LINKS

When the operator feed goes dark.

Jamming severs video and navigation links at the worst moment. Vektra engines run entirely on edge compute, holding recognition and tracking locally while the human decision is preserved and logged.

The generated Defense frame as a jammed video link delivers it: torn and frozen slices, dropped lines and noise Ground station · link lost
The same frame as held on the aircraft: a track, two howitzers and an infantry fighting vehicle at a treeline from directly above On board · frame held

Each pair is cut from the same generated frame: on the left, the crop as a detector working at a reduced input scale receives it; on the right, the same pixels at native resolution. The domain-shift pair sets two generated yards side by side, the thermal pair flips the palette of one frame, and the denied-link pair shows one frame as a jammed link delivers it and as it is held on board. They illustrate the conditions, not detection results.

Application intelligence

Recognition that fits
the next decision.

Connect visual results to asset records, inspection queues, operator review and downstream systems. Configure what the system recognizes and what its outputs mean to your teams.

  • 01Customer-defined classes and attributes
  • 02Context-aware application logic
  • 03Evidence linked to the originating observation
  • 04Deployment-specific performance tuning
Built around your challenge

What do you need
to recognize?

Start a project