CORE ARCHITECTURE · TECTUM
Perception is more than detection.
One forward compute pass builds a shared representation of the frame. Every capability reads from it. No second pass. No separate models.
01 · THE PROBLEM
The cascade is the cost.
EACH NEEDS ITS OWN TRAINING
PASS 1
Detector
PASS 2
Tracker
PASS 3
Re-ID network
PASS 4
Language model
// Multiple models to license, train, & maintain
// Each costs compute every frame
// Brittle hand-offs, and still no way to ask “which one”
Conventional stacks chain a detector, a tracker, a re-ID network, and a captioner, each licensed, trained, and maintained separately. Each burns its own pass of silicon. On edge power budgets, that math fails.
02 · ONE PASS
Encode once. Read everywhere.
Tectum computes a single rich representation per frame. Detection and localization, multi-object tracking, instance re-identification and cross-camera hand-off, concept gating, query-conditioned selection, on-demand scene description. All read from it. Marginal cost of a new capability: near zero.
DETECT
LOCALIZE
TRACK
RE-ID
HAND-OFF
GATE
SELECT
DESCRIBE
03 · ENROLLMENT
New target? Show it examples.
Tectum is not limited to a fixed catalog of classes. Operators enroll targets of interest from a handful of reference images, in the field, in minutes. From there the system labels its own training data: on public aerial benchmarks, zero-label enrollment reaches ~80% of fully supervised performance. Human annotations required: 0.
04 · THE GATE
Watch everything. Decode what matters.
Targets of interest trigger contextual enrichment stack only on contact while cutting enrichment time in up to half. Standby compute drops up to 23× versus describing every frame. Cost stays flat as scenes get dense.
23×
cheaper standby sensing
<50 ms
gate decision, per frame
0
additional encodes required
05 · ROBUSTNESS
Degrades gently where traditional object recognition models collapse.
Under motion blur and heavy compression, Tectum’s detection holds where conventionally trained detectors fall off. A domain shift from day to night or EO to IR is a 6-frame recalibration, not a retraining campaign. An on-board monitor watches the embedding stream and flags drift as it happens.
06 · DEPLOYMENT
Standalone or Hybrid. Both edge-first.
Tectum · Standalone
For most deployments, this is the whole system. Detection through description on one encode, real-time under 15 W on Jetson-class modules. The smallest models run on a Jetson Orin Nano.
ALONGSIDE EXISTING MODELS
Tectum · Hybrid
Where a trained detector already earns its keep, Tectum pairs with it. Your models stay, and enrollment, selection, and description ride the same shared encode.
// EDGE MODULE · <15 W REAL-TIME