Conventional RGB / fixed camera / private delivery
Fixed-camera sky video for hard-negative evaluation.
SkyStream Release 01 currently contains 9,698 delivery-ready clips and a 1.07-million-point event catalog collected from a continuously operated sky camera. The media is conventional RGB/H.264 with human corrections and frame-derived tracks, licensed privately for model evaluation. The same fixed rig can also continue collection against a written buyer brief.
9,698delivery-ready clips
60.375 hcorrected video duration
1,070,975trajectory points
378.144 GBdelivery video
01
Inventory
What is actually in the release
| Layer | Volume | Notes |
|---|---|---|
| Event video | 9,698 clips / 60.375 h | 9,688 source MP4s plus 10 lossless timing remuxes; H.264, no audio |
| Clip-linked catalog | 10,483 records | 785 records are metadata/trajectory-only and have no offered source clip |
| Primary resolutions | 6,512 at 3840×2160 3,169 at 2960×1668 | 9,689 of 9,698 clips are 2560×1440 or higher |
| Trajectory layer | 1,070,975 rows | 695,420 rows link to the current video inventory; relative frame time only |
| Human corrections | 2,636 events | 2,320 link to the current video inventory; provenance is preserved |
| Frozen benchmark | 2,340 images / 2,700 boxes | Reviewed COCO and YOLO formats; this is the Forge layer, not the whole corpus |
| Current linked media | 49,137 core/derived assets / 384.6 GB | Source-backed clips, keyframes, crops, overlays, stacks, and review derivatives |
02
A look at the data
Tracks and corrections, not just frames
Counts describe human-corrected events, not balanced training classes. Historical “meteorite” labels are presented conservatively as candidate streaks.
03
Fit and limits
Use it to find false alarms—not to claim object recall.
Good fit
- False-alert testing for conventional RGB sky cameras
- Hard-negative mining for birds, insects, cloud, foliage, aircraft, and sensor artifacts
- Domain-shift evaluation at an unseen fixed-camera site
- Track-level analysis for small-object and ambiguous-motion systems
- Sky-imager and atmospheric-camera false-trigger testing on conventional perspective RGB
- Optical space-monitoring hard-negative and image-cleanup evaluation
- Catalog-correlation review for individual satellite-like tracks
- Bird/bat monitoring, wind-energy vision, and optical C-UAS specificity work
Not represented
- No verified drone-positive class
- No paired irradiance, cloud fraction, meteorological calibration, or hemispheric/fisheye geometry
- No independently verified Starlink train, resident-space-object, launch, or meteor ground truth; the public satellite sample is an ephemeris correlation, not independent identification
- No native neuromorphic or event-camera stream
- No claim of geographic or camera diversity
- No species-level wildlife ground truth
- No public redistribution rights or unrestricted raw-media dump
Why one site can still be useful
A single fixed viewpoint is a controlled domain-shift test, not a claim of broad coverage. A model that has never seen this site either rejects its clutter or reveals where confidence, tracking, and alert logic fail.
04
Commercial access
Access and pricing
| Engagement | Scope | Price | Next step |
|---|---|---|---|
| Model audit | 200–500 stratified events; metrics, error gallery, findings memo | $2,500 fixed | Scope the audit ↗ |
| Evaluation subset | Buyer-specific clips, labels, formats, and internal evaluation rights | From $4,500 | Request terms ↗ |
| Full internal ML license | Nonexclusive training and evaluation rights; delivery scoped to buyer | From $15,000 | Discuss license ↗ |
| Continued fixed-rig capture | New outdoor RGB footage collected against an agreed condition, cadence, duration, and metadata brief | Quoted / milestone-funded | Send a brief ↗ |
Collection to specification
Custom capture extends the same fixed viewpoint; it does not add a second geography or guarantee rare objects or weather. Work begins only after the requested conditions, collection window, acceptance checks, privacy treatment, rights, milestones, and funding are agreed.
Fixed audit path
The audit, in four steps.
A technical fit note starts the conversation. No source corpus, buyer model, credentials, or commercial rights move at this stage.
- 01Fit
Confirm camera modality, target classes, false-alert cost, and useful outputs.
- 02Scope
Agree the 200–500-event cohort, success measures, security route, and written terms.
- 03Run
Evaluate one buyer model or API on the frozen cohort after access and payment are settled.
- 04Return
Deliver per-condition metrics, an error gallery, and a short findings memo.
Start with four details: camera modality, target classes, the costliest false alerts, and whether model access is available by container, offline executable, or API. Never email credentials or proprietary model files.
Start with four details ↗05
Provenance and privacy
Source-controlled, with operational details withheld
Offered media and first-party annotations were captured or created under the supplier’s control. Exact timestamps, exact location, internal paths, credentials, and identifying content are withheld or redacted as applicable.
Third-party ADS-B, weather, satellite, map, and external-imagery fields are excluded from the commercial package. Original-media access and redistribution rights are controlled by buyer-specific terms.
Redacted evaluation sample / v3.2
Inspect the files before starting a sales call.
Fourteen H.264 clips, paired clean and annotated keyframes, a manifest, 128 normalized trajectory points, validation results, and file hashes. Camera overlays and source metadata are removed. The sample grants no training, redistribution, publication, or production-use rights.
Play ten clips in the browser No account / no download required Download evaluation kit v3.2 ZIP / 3.2 MB / 50 files
SHA-2567c2f060dd602adda0dbedbf02426e1ad8452a70519999a92f386e73a537a27f1
Technical review
Need a buyer-specific slice or model audit?
Edmund Olsen / SkyStream
edmundolsen7@icloud.com ↗