Docs
Current software

Public product routes

Five public routes accept operator inputs or expose a prepared example without requiring an account. This page records what each route does and where its operating boundary stops.

Praxis Mission Studio
Run a short video through detection, identity tracking, one rule, and retained evidence.
Praxis Memory
Build a local multi-view case or inspect the documented prepared example.
Praxis World
Open a local point cloud and camera path, or inspect the bundled reconstruction example.
Praxis Watch Areas
Create a reusable fixed-camera objective or inspect the registered aerial example.
Praxis Model Arena
Score local detection outputs against ground truth and inspect exact error frames.

Praxis Mission Studio#

Mission Studio accepts a short operator video or its bundled overhead example. A mission selects a subject family and either watches a marked area for entry, exit, or dwell events, or follows one selected subject. The hosted CPU worker produces a playable annotated result, event timestamps, source frames, track observations, checksums, and a downloadable Evidence Pack.

Mission Studio boundary

The public worker queues accepted missions and processes one at a time with per-workspace and global hourly limits. Videos are limited to 60 seconds and 80 MB. Projects control retention, versioned objectives, and run history. API keys belong to the owning browser workspace. Owner-authenticated artifacts use expiring URLs, and read-only shares can be revoked.

Praxis Memory#

Memory accepts two to four local video files in the browser. The operator names each view and sets which source time belongs at the current case time. One transport then plays, pauses, and scrubs every available recording on the shared clock.

A link records that the operator confirmed the same subject or event across at least two views. The portable JSON contains filenames, durations, offsets, link times, and notes. Lightweight case metadata can also be saved in the current browser.

Memory boundary

Memory does not upload local video, estimate source offsets, associate people automatically, or perform biometric identification. Browsers do not preserve access to local files after a reload, so reopening saved metadata requires the operator to reselect each recording.

Prepared example

The Example tab uses Scenario 2 from the public NANCY multi-perspective traffic recording. Its drone, roadside, and vehicle derivatives were aligned by hand around the staged event. One participant was seeded manually in each view and followed within that view with OpenCV CSRT. The fixed OSNet x0.25 scores compare three prepared crops at the 2.0 second case moment and are not current-time inference.

Praxis World#

World opens operator-selected PLY, XYZ, TXT, PTS, and CSV point clouds entirely in the browser. An optional CSV adds camera positions. The operator can choose Y-up or Z-up, auto-frame the geometry, orbit the scene, and return to the bundled example.

Local import boundary

World renders an existing point cloud. It does not reconstruct a scene from raw imagery. Imported file bytes are not uploaded or saved. Only filenames, point counts, file sizes, and view settings may remain in local browser storage.

Prepared example

World uses Open Forest Observatory mission 000173. The published source mesh contains 4,217,056 vertices and 8,430,086 faces. Praxis ships a uniformly sampled and transformed browser derivative with 263,566 colored points, plus 800 recorded camera positions.

The four inset images come from the published source. They cycle independently and are not presented as matches to the selected position. The capture records, source page, license, and derivative restriction remain visible together.

Prepared example boundary

Praxis did not create the published reconstruction. The interface displays recorded positions, not camera orientation, view direction, or frusta. Its coverage replay assigns each displayed point to the nearest recorded position. That ordering is an interface aid, not the chronology of the original reconstruction. Mission 000173 is used for evaluation and visualization only, not model training.

Praxis Watch Areas#

Watch Areas starts a Mission Studio area objective for operator-selected fixed-camera footage. The operator draws a polygon, chooses entry, exit, or dwell, saves the objective to a project, and can apply it to later clips. Completed runs retain timestamped matches and their source frames.

Watch Areas boundary

User-created Watch Areas currently refer to fixed-camera framing. The aerial example uses checked-in registered geometry and track data from one recorded flight. It demonstrates the moving-camera method but does not run that registration pipeline in the browser.

Praxis Model Arena#

Arena accepts one ground-truth JSON or CSV and two to four detection-result files. It applies class-aware one-to-one matching at configurable confidence and IoU thresholds, reports precision, recall, F1, TP, FP, and FN, and opens the frame geometry behind each error. The complete comparison exports as deterministic JSON.

Model Arena boundary

Arena evaluates result files locally. It does not execute arbitrary model weights. The prepared tracker replay remains a separate example and its continuity proxy is not labeled as ground-truth accuracy.

Source records#

Mission Studio and saved Watch Areas use the hosted worker. Memory, World, and Model Arena process operator-selected files in the browser. None of these routes represents a hardware flight test, deployment benchmark, or field validation.