RenderWorks
Discuss a project

VISION AI DATA INFRASTRUCTURE

Beyond detection.Understand context.

Real-world event datasets for training and evaluating Physical Security AI.

RenderWorks is a Vision AI data infrastructure company that designs, captures, structures and validates real-world datasets for AI training and evaluation, with a current focus on Physical Security AI.

Discuss a Dataset Project

01 / EXPERIENCE

Real-world experience,measured in data.

RenderWorks and its core team have worked across controlled scenario capture, multi-camera footage, multimodal sources, and structured annotation workflows.

10,000+VIDEO CLIPS

cumulative data-production experience

550+VIDEO HOURS

real-world capture experience

4AI CCTV DATA PROJECTS

RenderWorks + core-team experience

EXPERIENCE ACROSS

  • CONTROLLED EVENT SCENARIOS
  • MULTI-CAMERA VIDEO
  • RGB · IR · DEPTH
  • TEMPORAL · BBOX · POSE · ROI

Figures combine RenderWorks project experience and prior experience of the core team. Detailed project references are shared where disclosure permits.

02 / FAST REAL-WORLD BUILD

3,000 real-world clips,built on a practical baseline.

Reference configuration: 2+ cameras, controlled real-world capture, clip editing, metadata, and scoped GT / Map / geometric labels. Additional sensors, dense annotations, augmentation, and synthetic data are extended by project scope.

UP TO 200EVENT EXECUTIONS / DAY

12-hour day + night controlled capture

2+CAMERAS

baseline multi-camera configuration

3,000VIDEO CLIPS

reference real-world dataset scale

≤ 2 WEEKSTARGET BUILD WINDOW

capture → edit → scoped annotation → QA

Reference production target, not a fixed SLA. Timing and output vary with location access, event complexity, annotation density, and additional sensor requirements.

03 / FROM FOOTAGE TO STRUCTURE

Structure what happened,not just what appeared.

One real-world event becomes progressively more useful to AI as temporal context, roles, objects, interactions, and geometry are structured around the source footage.

PRIMARY

LAPTOP THEFT

Context establishes role before the removal event is interpreted.

OPEN BENCHMARK SAMPLE
LAPTOP THEFT
  1. 01
    CONTEXT

    Read movement and roles before the short event.

  2. 02
    ROLES + MOVEMENT

    Connect accumulated movement context across owner and actor.

  3. 03
    OBJECT

    Isolate the laptop as an object.

  4. 04
    STRUCTURE

    Use pose and scene boundaries to narrow the interaction.

  5. 05
    SPACE

    Derive spatial relationships with ESTIMATED DEPTH.

  6. 06
    STATE CHANGE

    Confirm the laptop's disappearance through the before-and-after state change.

The same structuring method follows a different interaction: an object left behind.

SECONDARY

OBJECT ABANDONMENT

Hand and object evidence make the state transition visible around the event clip.

OPEN BENCHMARK SAMPLE
OBJECT ABANDONMENT
  1. 01
    EVENT

    Find the interaction within the short action interval.

  2. 02
    HAND + OBJECT

    Narrow the view to the hand ROI and bag object.

  3. 03
    HAND / OBJECT SEPARATION

    Connect the change as the hand and object move apart.

  4. 04
    STRUCTURE

    Use pose and scene boundaries as bounded evidence.

  5. 05
    SPACE

    Derive spatial relationships with ESTIMATED DEPTH.

  6. 06
    STATE CHANGE

    Confirm the bag was left behind through the before-and-after state change.

VIDEO + ANNOTATION + METADATA + PROVENANCE + QAMODEL-READY DATA

Open benchmark footage used for demonstration. The displayed evidence is the result of annotation work performed by RenderWorks on the original footage.

04 / PRODUCTION PIPELINE

Capture. Inspect.Edit. Validate.

RenderWorks operates data production as one repeatable workflow from controlled field capture to dataset QA and delivery readiness.

  1. 01
    CAPTURE

    Reproduce the required event in the real world.

    Scenario design · controlled environment · actors · camera plan · day / night conditions

    OUTPUTSOURCE FOOTAGE
  2. 02
    INSPECT

    Verify that the inputs can support the intended dataset.

    Media integrity · camera relation · sync · metadata · scenario correspondence

    OUTPUTVERIFIED INPUTS
  3. 03
    EDIT

    Turn source footage into structured event data.

    Event clips · temporal GT · BBox · Pose · ROI · segmentation · geometry · metadata

    OUTPUTSTRUCTURED DATASET
  4. 04
    VALIDATE

    Check the final package before evaluation or delivery.

    File / metadata consistency · GT relation · visual review · QA · evaluation readiness

    OUTPUTDELIVERY-READY DATA

Internal software supports the pipeline, but the production workflow—not the application UI—is the public product evidence.

05 / HUMAN-IN-THE-LOOP

HUMAN-IN-THE-LOOP DATA PRODUCTION

Automate repetition.Review ambiguity.

Internal tooling reduces repetitive production work while ambiguous evidence remains subject to explicit human review.

AUTOMATERepeatable production work

Sync, clipping, structured inputs, proposal generation, and deterministic checks are supported by internal tools.

REVIEWAmbiguous visual evidence

Occlusion, role ambiguity, event boundaries, multi-person scenes, and model-assisted proposals are reviewed before final GT is accepted.

06 / RW DATASET

OPEN DATASET IN PREPARATION

RW DatasetComing Soon.

An open real-world event dataset for Physical Security AI research is currently in preparation.

Designed for reuse across broader research and public benchmarking. Release scope, license, and annotation details will be announced with the dataset.

Discuss a Dataset Project