For Physical AI
Selected objects and scene features.
- Objects
- Drone; selected bird at 25–25.5 s
- Infrastructure
- Transmission pylon crossarm
- Handoff
- Frame-linked annotation coordinates
See the data.
Explore the detail.
See how ground, air and maritime footage becomes object labels, scene features and structured observations. Source and review notes distinguish saved annotations from illustrative drafts.
Ground robotics, in context.
A tracked machine traverses an indoor obstacle course. Selected object annotations connect the machine to its surroundings.
Video could not load. View the annotated frame or .
Observation & limits
- Observed
- A tracked machine changes position and orientation in an indoor gravel and log course.
- Unknown
- Control mode, autonomy and performance outside this demonstration.
Source & annotation method
Editorial boxes on sampled frames of real footage, with linear interpolation, prepared to illustrate a possible project output. No validated tracking, condition assessment or production export is claimed.
Read the annotation & observation recordDescribe what is visible.
Visible hull deformation can be recorded alongside vessel annotations, while keeping cause and operational condition unresolved.
Video could not load. View the annotated frame or .
Observation & limits
- Observed
- An irregular hull opening or deformation, partly obscured by covering and an adjacent vessel.
- Unknown
- Cause, date, severity, independently confirmed identity, repair status and seaworthiness.
Source & annotation method
Editorial boxes on sampled frames of real footage, with linear interpolation, prepared to illustrate a possible project output. No validated tracking, condition assessment or production export is claimed.
Read the annotation & observation recordObjects, labelled
in context.
Vessel masks and object labels make this recording useful for a perception dataset. View the annotated frame or play the short excerpt.
Saved annotation draftVessels on open water
Silent · playback starts on request
Video could not load. The annotated frame and record remain available above.
What you are looking at
An overhead recording of vessels on open water. The saved annotation adds vessel masks and object labels. This is frame 540, at 18.0 seconds of the source recording.
The four-second annotated excerpt starts at 16.5 seconds in the source recording.
Annotation example carried forward from the FORTMOSA expo materials. Source recording: cg_interdiction.
Review status matters
Reviewed seeds with machine propagation. The saved draft is structurally validated and has sampled visual review. It is not fully human-approved ground truth.
A delivery would agree label definitions, review coverage, metadata and export format for your task.
Inspect the matched annotation record
- Source time
- 18.0 seconds
- Source frame
- 540 · 30 fps
- Source dimensions
- 1280 × 720 px
- Example track
- 1
- Export class
- boat
- Box (x₁, y₁, x₂, y₂)
- 577, 0, 1179, 209 px
- Render note
- The image retains “Ship 01”; the canonical exported class is boat.
Perception in practice.
Different scenes need different checks. Explore masks, object tracks and human corrections in real footage. All clips are silent and play on request.
Landing in dust
Follow an object as its appearance changes. The mask marks the drone; the track links its position across frames.
Video could not load. View the annotated frame or .
Draft annotation · 2.5-second clip
Source & review notes
One drone is annotated. Dust obscures fine edges; background structures are excluded. This is a research draft, not accepted training ground truth.
Collector 1014978 · “173rd Airborne Gets Comfortable With Drones”, segment 23, collected via Funker530.
Tracking on water
Keep a moving object distinct from reflections and wake. The saved mask and track show what the annotation follows.
Video could not load. View the annotated frame or .
Draft · sampled review · 4-second clip
Source & review notes
Reviewed seeds with machine propagation, structural validation and sampled visual review. The complete track has not been human-approved.
Sochi port recording · FORTMOSA expo annotation bundle. Excerpt starts at 4.5 seconds in the source.
Where human review helps
Small objects are easy to miss. Pink boxes show tracker output; amber boxes show corrections made by an annotator.
Video could not load. View the annotated frame or .
Reviewed tracks · 10-second excerpt
Source & review notes
The accompanying export reports every track reviewed and 2,399 of 5,287 boxes adjusted across the full flight test. These are review counts, not an accuracy score.
Collector flight-test export · source frames 3510–3809 (1:57–2:07). The website has not independently audited every box.
Several objects, separate tracks
Keep separate identities as objects move and overlap. Each colour belongs to a saved object track.
Video could not load. View the annotated frame or .
Draft annotation · 2.8-second clip
Source & review notes
Eight drones are annotated within declared windows. Close overlap, exits and missing masks need further review; smoke and debris are excluded.
Collector 1021452 · “Skynex Anti-Air System Smokes Drones”, segment 00, collected via Funker530. Rheinmetall watermark retained.
Eight drones.
Shown twice.
A recording shows a group of drones, then repeats the sequence in slow motion. Recognising the replay helps avoid counting the same objects again.
First appearance
Eight drones in the opening sequence.
Video could not load. View the annotated frame or .
Slow-motion replay
The same group appears again later in the recording.
Video could not load. View the annotated frame or .
Two appearances of the same group
Keep the relationship with the footage. A new appearance or annotation ID does not necessarily mean a new object.
What connects these excerpts?
Both come from collector clip 1021453. Its annotation notes identify eight drones shown twice, with 16 appearance IDs across the opening and replay. The source itself labels the replay “Slow Motion”.
“Skynex Anti-Air System Smokes Drones”, segment 01, collected via Funker530. Existing Rheinmetall watermark and annotation colours retained.
What has been checked?
The excerpts use saved draft annotations with partial object coverage. Overlap and missing masks still need human review.
The relationship is explained here from the footage and its notes. This is not an automated deduplication result or a verified event timeline.
Explore a source-grouping workflowControlled imagery.
A matched reference.
Can a generated record be inspected against the frame it describes? This synthetic example pairs one camera image with its generated reference. The target is only a few pixels across; the enlarged detail makes it easier to inspect.
Object present · reference position marked640 × 480 px
No additional detail generated.
The record belongs
to this frame.
- Dataset example
- clip_0000 · frame 0
- Projected position
- x 353.06 · y 258.07 px
- Rendered centroid
- x 352.71 · y 256.95 px
- Reference
- Generated EO ground truth
These coordinates describe the object-present frame. The absent view is its matched background, not another labelled target.
Supplied synthetic research material: Axair / detection research inputs. Sweep provider recorded as Axair GeoEngine.
Why use synthetic examples?
Controlled scenes can add conditions that are difficult to collect. Keep synthetic and real data distinguishable, and check whether the generated conditions serve the task.
What this does not establish
This saved image and reference demonstrate a matched record. They do not establish calibrated sensors, a live LOKE simulator integration or performance on real footage. Partner support and exports require a defined scope.
One observation.
Several publications.
Follow one recording through its reposts, then compare a second viewpoint and a claimed update. Select a recording to inspect what is known and what still needs checking.
Fictional scenario: an access route is reported blocked.
A civilian recording is reposted twice, a fixed camera offers another viewpoint, and a later drone post claims the route has reopened. Explore how to keep these records and their unresolved claims separate. This is a workflow illustration, not incident evidence.
Explore the observations
A fictional access-route disruption
Publication order is not occurrence time.
Repeated footage, preserved sources.
One recording · three publicationsPublic post A, a reshare B and a cropped excerpt C appear to carry the same recording. Grouping them avoids counting repeated circulation as independent corroboration.
- Public post AEarliest located publication · creator unknown
- Reshared post BSame footage · publication history retained
- Cropped excerpt CLikely edited copy · relationship needs review
Grouped for reviewRecording A
Still open for review
The caption's location and recording time are unverified. The earliest located post is not necessarily the original recording.
Sources can look very different
A pilot could work with agreed fixed-camera recordings, drone footage and civilian media shared publicly during a crisis. Preserve each publication's history even when recordings are grouped.
Captions, dates and claimed locations are assertions to assess. Edited copies and independent viewpoints need separate relationships.
The proposed handoff
A source-linked timeline, grouped observations, review decisions and unresolved questions. Begin with one incident type, one context and a defined time window.
Occurrence time, publication time and ingestion time remain distinct. A detection or a source's report is not a confirmed incident.
Discuss a retrospective pilot