Oceanova / Marine Intelligence 2.0.1 Early release

01 / Ocean intelligence software

BioSync

From underwater video
to ocean intelligence.

Get the Windows preview Explore the software

Live video / Local AI / Geospatial reviewDesigned & developed in New Zealand

02 / A new perspective

See beyond
the footage.

BioSync brings video acquisition, experimental AI-assisted observation, species review and mapping into a single desktop workspace. Built for marine teams working with real cameras, survey footage and imperfect underwater conditions.

Live & recorded videoAI-assisted reviewOnline / offline GISLocal-first processing

BioSync / Library

BioSync desktop software library interface screenshot

Observation library01 / Screen

03 / Library & review

One workspace.
Better context.

Bring candidate observations, reference imagery and reviewed records together. Return to what matters without losing the surrounding video and field context.

BioSync / Annotations

BioSync marine video annotation interface screenshot

Human-in-the-loop02 / Screen

04 / Human-in-the-loop

AI suggests.
You decide.

Review detections, correct uncertain labels and record annotations for later analysis. BioSync is designed to support marine professionals, with decisions and identification remaining under human review.

05 / Deep-water workflows

Look closer.
Review deeper.

Explore the annotation workspace across underwater footage, from camera survey imagery to more challenging subsea scenes.

BioSync / Deep-water annotations

BioSync underwater deep-sea annotation screenshot

Deep-water video review03 / Screen

BioSync / Capabilities

Designed for the field.

Capture

Work with supported cameras, RTSP streams and recorded video.

Analyse

Use local experimental vision models to generate candidates for review.

Validate

Review annotations and keep uncertainty visible rather than hiding it.

Map

Place observations in geospatial context and export reviewed data.

Next / Technical release

Ready to explore
the software?

Specifications & downloads

BioSync 2.0.1 is an early community trial. AI predictions, species suggestions and counts are experimental and require human review; this is not a validated biological survey instrument.

06 / BioSync technical

Native.
Local.
Reviewable.

Windows 11 x64 / Community preview

BioSync is a native desktop marine workstation designed around local video, AI-assisted review and geospatial data, without requiring a cloud inference service.

System specifications.

BioSync / v2.0.1

Workstation & compute

01 / HARDWARE

Operating systemWindows 11 / x64

System memory16 GB base / 32 GB recommended

Marine AI GPUNVIDIA CUDA / 16 GB VRAM recommended

GPU driverCompatible with CUDA 12.8 runtime

Available disk space40 GB + additional survey storage

AI model downloadApproximately 15 GB, separate

The NVIDIA configuration reflects the recommended trial setup, not a guarantee of compatibility on every GPU. Python, Qt and the application runtime ship with the installer.

Software & data

02 / WORKFLOW

Desktop frameworkQt 6 / QML / PySide6

Video inputsSupported USB / UVC / RTSP / local files

AnalysisLocal AI suggestions + human review

GISOnline / offline maps

Local dataSQLite observations

ExportCSV / GeoJSON

Internet is needed for initial model downloads and online resources. Installed local media and inference workflows can be used offline, subject to device compatibility.

One connected workflow.

Input / AI / Review / Export

01 / INPUT

Video

Connect a supported camera, stream or recorded media file.

02 / PROCESS

Local AI

Generate experimental candidates with optional local models.

03 / REVIEW

Annotate

Check suggestions, correct labels and keep reviewed records.

04 / OUTPUT

Map & export

Inspect spatial context and export observations for analysis.

07 / Software release

Get BioSync.

Register your interest to evaluate the Windows community preview, then download the four required installer files directly from Oceanova's official GitHub Release.

2.0.1Windows x64 / Early release / ~3.1 GB

Windows early trial

BioSync 2.0.1 is available as a four-part setup package. Submit the short trial registration below to display the files. This form notifies Oceanova of your intended evaluation.

Windows 11x64Preview3.1 GB total

Already registered? Resume a download on GitHub ↗

Installation & release notes

Follow these steps for the Windows preview. Each numbered BIN is a data component read by the installer; do not try to open BIN files individually.

Submit the preview form and download the setup EXE plus all three numbered BIN files.
Save every installer file together in a single folder. Keep each original filename unchanged.
Run the setup EXE. The desktop shortcut is enabled by default.
For full AI, open Tools → Installation / updates / local AI and download the additional models (about 15 GB).

Read me first ↗ Full installation guide ↗ Verify checksums ↗

What is included in 2.0.1?

Native desktop workstation, supported video inputs, experimental AI-assisted species review, local records and GIS tools. Development-only demos and experimental tracking modes are not part of this public installer.

Why is the installer unsigned?

Code-signing is pending for this preview. Windows may show a publisher warning. Review the download source and supplied SHA-256 hashes. Never disable Microsoft Defender, SmartScreen or your firewall to install it.

Are Linux and macOS downloads available?

No public Linux binary or macOS package is included in the 2.0.1 preview. Linux validation is in progress; supported files will be listed on an official release when available.

Feedback / Early trial

Help shape the next release.

Found a reproducible issue or camera incompatibility? Please include your OS, GPU, input type and steps to reproduce. Do not upload private survey footage or camera credentials.

Report an issue on GitHub ↗

Data & integrity

Local by design.

BioSync stores user data under your local Oceanova BioSync profile. Back up important work before installing an updated build. Model downloads are separate; detection outputs remain unverified suggestions for human review.

08 / System architecture OCEANOVA / BIOSYNC / NATIVE MARINE COMPUTE

ENGINEERED
FOR THE
DEEP.

Computer vision is easy to demonstrate. Scientific evidence is harder to engineer. BioSync separates acquisition, detection, tracking, biological inference and human verification rather than reducing marine observations to a single AI label.

01 / UIQt 6 / QMLNative video and operator interface

02 / AIPyTorch / CUDAOn-device model inference

03 / STORESQLiteEvidence and reviewer provenance

04 / GISWGS 84 / MBTilesSpatial context, online or offline

08.1 / Processing topology

Not one model.
One instrument.

The video renderer and inference worker have different jobs and different timing. Biological ranking and multimodal interpretation are asynchronous stages, not a blocking operation on every frame.

PIPELINE / LOCAL WORKSTATIONDECOUPLED EXECUTION

VIDEO ACQUISITION + NATIVE DISPLAY

USB / UVC · RTSP · local media · Qt Multimedia / FFmpeg

UI / VIDEO

↓

FRAME SAMPLING + PROCESS BOUNDARY

Frame resize · JPEG transfer · bounded IPC queue

ASYNCHRONOUS

↓

DETECTION + TEMPORAL ASSOCIATION

OWLv2 proposals · IoU matching · pyramidal Lucas-Kanade flow

TRACK CONTEXT

↓

INDEPENDENT, NON-BLOCKING ANALYSIS PATHS

04A

VISUAL + BIOLOGICAL EVIDENCE

DINOv2 embeddings · BioCLIP 2.5 ranking · references

SIMILARITY

04B

MULTIMODAL INTERPRETATION

Qwen3-VL · diagnostic features · uncertainty

LANGUAGE

↓

OBSERVATION RECORD + HUMAN REVIEW

Source · crop · model evidence · verification · SQLite / GIS export

TRACEABLE

Conceptual processing topology, not a fixed per-frame execution sequence. Detection and initial evidence may be persisted before asynchronous biological interpretations complete.

08.2 / Neural stack

Specialists, not
one black box.

Model families perform complementary operations. An uncertain result at one stage cannot be repaired simply by generating a more confident-sounding answer downstream.

OWLv2

google/owlv2-base-patch16-ensemble

Open-vocabulary localisation. Text-conditioned proposals identify candidate image regions. The detector outputs normalised boxes and scores; overlaps are filtered before track association.

Failure boundaryMay miss cryptic organisms or produce false positives in cluttered, turbid or low-contrast imagery. Detection score is not species confidence.

DINOv2

facebook/dinov2-small

Visual embedding and retrieval. Candidate crops are represented by normalised feature vectors for similarity lookup against previously stored imagery and references.

Failure boundaryVisually similar taxa can cluster together. Nearest-neighbour appearance does not establish taxonomic identity.

BioCLIP 2.5

imageomics/bioclip-2.5-vith14

Biology-oriented candidate ranking. Image/text embeddings rank biological names and reference examples; reference evidence and coarse group constraints can inform ordering.

Failure boundarySimilarity ranks are uncalibrated. A taxon existing in the library is not evidence that its field classifier has been validated.

Qwen3-VL

Qwen/Qwen3-VL-4B-Instruct

Evidence-assisted interpretation. Selected imagery and candidate context support structured descriptions, visual feature comparison and explicitly uncertain identity suggestions.

Failure boundaryA plausible narrative is not a measurement. Hallucinated diagnostic features, unsupported taxa and incomplete interpretations require rejection or review.

Geometry / associationIoU(A,B) = |A ∩ B| / |A ∪ B|

Spatial overlap helps match detector proposals to existing tracks; it does not solve identity through long occlusions.

Feature retrievalsim(a,b) = (a · b) / (||a|| ||b||)

Cosine similarity compares embedding directions, not calibrated likelihoods of biological membership.

Coordinate semanticsbbox = [x, y, w, h], 0..1

Normalised image-frame geometry is separate from geodetic coordinates, depth and physical range.

09 / Developer & research documentation

LOOK
CLOSER.

For engineers, integration teams and researchers who care about process boundaries, feature spaces, track continuity, positioning provenance and validation methodology. Open any module below for implementation notes.

01Inference scheduling, IPC and failure isolation

The Qt workstation decodes and renders video independently of the Python inference worker. Selected frames are resized, encoded and submitted through a bounded multiprocessing queue so the application does not accumulate an unbounded backlog when the GPU is busy.

The current desktop implementation starts its inference adapter in a spawned process. The worker validates returned detections and attaches the source generation, capture time and elapsed inference latency. The interface discards results from superseded sessions and rejects stale outputs.

Engineering implication: source frame rate, analysis submission rate, detector update cadence and Qwen interpretation latency are separate quantities. The system prioritises an interactive operator interface over processing every frame synchronously.

Logical scheduler

QVideoFrame
  | resize + encode
  v
request_queue [bounded]
  |  separate process
  v
GPU inference / ranking
  |
result_queue + timestamps
  v
accept if generation valid
and result age acceptable

Pseudocode, not a stable public plugin API. Worker errors do not require video playback to stop.

02Object detection, optical flow and track association

OWLv2 provides fresh spatial proposals. Between these detection passes, pyramidal Lucas-Kanade optical flow estimates displacements from track-region feature points, using a median motion vector to advance the candidate box.

New detector proposals are matched to recent tracks by Intersection-over-Union, while missing or invalid optical-flow evidence can cause tracks to expire. A change of session, analysis mode or frame geometry resets state.

Track IDs represent short-lived visual associations. Turbidity, overlapping animals, rapid motion and occlusion can cause lost tracks and identity switches. Neither optical flow nor box matching establishes an individual-animal census.

Tracking concepts

I(x, y, t) ~= I(x+u, y+v, t+dt)

feature points -> LK flow
median(dx,dy) -> box shift
fresh proposals -> IoU match
unmatched -> new track
stale / invalid -> expire

Short-term motion estimation. Not appearance-based long-term animal re-identification.

03Reference embeddings and biological candidate fusion

DINOv2 descriptors support visual retrieval, while BioCLIP 2.5 uses biology-oriented image and text embeddings to organise candidate taxa. The biological ranker can combine text similarity with visual reference support and detector-informed coarse taxonomic constraints.

Ranking happens separately from the slower Qwen interpretation path. Selected evidence and reference examples can then inform visual descriptions or tentative identities, with a reconciliation stage checking whether an asserted taxon is actually supported.

No single score is presented as a calibrated posterior probability. Model agreement can still be wrong if inputs are out of distribution or two taxa share similar visible features.

Evidence categories

detector_score  = box proposal
embedding_sim   = visual proximity
ranking_score   = candidate ordering
LLM_description = interpretation
review_state    = human decision

ranking_score != P(taxon | image)

Outputs are evidence of different kinds, not interchangeable confidence values.

04Multi-view consistency and human-in-the-loop learning

The development architecture includes bounded temporal evidence: multiple sufficiently separated candidate views can be compared for agreement, while near-duplicate crops are prevented from masquerading as independent evidence.

This supports consistency checks but is deliberately conservative. Three repeated wrong suggestions are still wrong; agreement does not convert a machine prediction into a scientific verification.

Reviewed examples support future dataset curation, evaluation and specialist model development. A corrected label does not automatically retrain the deployed foundation models. Evaluated specialist models must meet their own activation and compatibility criteria.

Evidence policy

candidate crop + timestamp
    |
separation + duplicate check
    |
compare ranked hypotheses
    |
provisional consistency signal
    |
HUMAN REVIEW remains required

Current code uses a bounded multi-view evidence window. Temporal consensus is not taxonomic proof.

05Observation schema, source provenance and subsea positioning

SQLite observation records can retain source identifiers, timestamps, provisional labels, detector scores, normalised boxes, image crop paths, embeddings, reference matches, model predictions and reviewer decisions. Export paths support CSV and GeoJSON.

The GIS layer handles WGS 84 reference data, supported MBTiles and XYZ tile sources, and local spatial records. Surface vessel GPS is not an underwater target measurement. Suitable underwater position fixes may be retained as explicit survey-position proxies with their accuracy and provenance.

Unlocated observations should remain unlocated, including in geospatial exports. Camera position, depth, attitude and target range would be needed for a stronger claim of actual organism geolocation.

Conceptual observation

{
  "source": "RTSP / recording",
  "track_id": "session-local",
  "bbox": [x, y, w, h],
  "model_evidence": {...},
  "review": "unverified",
  "position_basis": "surface | underwater",
  "target_location": null
}

Illustrative field structure; not a literal SQLite schema or supported import API.

06Field video, offline maps and network boundaries

The source input layer includes native camera access, file playback and compatible RTSP streams, with an OpenCV/FFmpeg compatibility path for supported decoders. Device discovery does not guarantee that its media stream or codec is actually usable.

Local inference can operate offline after the correct model assets have been installed. Offline maps require map packages or cached coverage to be prepared in advance. Provider licences govern whether particular tiles may be stored or redistributed.

Compatibility trials should test actual ROV/camera hardware, authentication, codec negotiation, extended running, stream loss and recovery. Network access is still used for initial model downloads, some reference resources, online maps and user-requested update checks.

Deployment boundary

CAMERA / FIELD LAN
  RTSP / TCP, USB, local video

WORKSTATION
  Qt media + local CUDA models
  SQLite + cached GIS resources

ONLINE WHEN REQUESTED
  models / map tiles / updates

Offline inference does not imply every external data source is available offline.

10 / Test methodology

Measure the system.
Not the demo.

Evidence of field performance means testing on independent survey recordings, not just the images used for model selection or reference indexing. BioSync is an experimental workstation; biological accuracy and performance depend on hardware, imagery and operating conditions.

Evaluation should separate object detection, track quality, biological ranking, open-set rejection and compute latency. We will not imply that an impressive visual overlay is itself a validated ecological measurement.

Suggested validation outputs / Not measured claims

DetectorPrecision, recall, false positives, missed targets, class/size sensitivity

TrackingIdentity switches, fragmentation, recovery under occlusion or camera motion

TaxonomyTop-k ranks, per-taxon errors, unknown rejection, out-of-distribution behaviour

ComputeLatency distribution, analysed FPS, video FPS, sustained VRAM allocation

EnvironmentTurbidity, lighting, depth, viewpoint, resolution and camera variation

10 / Field Guide BIOSYNC 2.0.1 / WINDOWS PREVIEW

FROM FIRST FRAME
TO FIELD EVIDENCE.

A practical operator's walkthrough of the BioSync workstation: connect a source, prepare a map, analyse imagery, review candidates and export your observations.

Built for fieldwork. The guide follows the current BioSync 2.0.1 preview interface. Every screenshot opens at full resolution so you can inspect the controls before using them.

BioSync 2.0.1 screenshot: 00 / Workstation overview — The main window brings the video, survey GIS, detection panel and local status together.Open full size ↗

Preparation

Before you begin.

01 / Install

Use the Windows 2.0.1 installer. Start BioSync from the desktop shortcut.

02 / Download models

Open Tools → Installation / updates / local AI... and select Download / repair local AI models for full marine AI.

03 / Prepare a source

Have a USB camera, compatible RTSP URL, or sample photo/video ready. A live connection is not required to explore recorded media.

Chapter A / Input

Connect the survey.

CAMERA → VIDEO → GIS

01 / 06 OPERATING PROCEDURE

Bring in a video source.

Start with the camera or recording you already have. BioSync can work with local footage as well as compatible live devices and network video streams.

Select Video → Connect source... or choose Source... in the top toolbar.
For a USB camera, select the discovered device and press Connect. For a network camera, enter the device's RTSP URL and select the decoder where required.
To work from recorded material, use Open... or File → Open photograph / video.... Confirm the video appears and the status changes to LIVE.

Connection check: Camera discovery is not a guarantee of stream compatibility. If a feed will not connect, verify the URL, codec and transport, or try the compatibility decoder.

BioSync 2.0.1 screenshot: 01 / Video source dialog — Choose live or recorded input; use the manual source field when auto-discovery is unavailable.Open full size ↗

Chapter B / Geospatial setup

Give the footage context.

BASEMAP → SURVEY LAYERS

02 / 06 OPERATING PROCEDURE

Choose a survey basemap.

Use the integrated GIS to orient your fieldwork against coastline, place names and available map layers.

Open Survey → Basemap / offline downloads..., or select Maps... above the survey map.
Choose an available map source. Pan and zoom the map to your area of interest; use Offline only when working without network basemap access.
For offline fieldwork, set an area name and zoom range, then choose Download / resume area where the provider permits it.

Field note: Offline area preparation has size and tile limits. Verify the saved map at the intended zoom levels before travelling offshore.

BioSync 2.0.1 screenshot: 02 / Basemap and offline downloads — Work with the built-in basemap and prepare a bounded offline survey area.Open full size ↗

03 / 06 OPERATING PROCEDURE

Add custom imagery or map tiles.

Bring in a site-specific aerial or raster source where your mapping provider supports XYZ tile access.

Within the map settings, open Custom imagery / XYZ source.
Enter a Source name, a permitted HTTPS XYZ tile template such as /{z}/{x}/{y}.png, required attribution and supported maximum zoom.
Choose Save and use source. For existing offline datasets, use Open MBTiles... or XYZ folder... instead.

Licensing matters: Only enable offline downloading when your imagery provider explicitly permits it. An online tile URL does not imply redistribution or offline rights.

BioSync 2.0.1 screenshot: 03 / Add a basemap source — Configure custom XYZ map imagery with its attribution and provider permissions.Open full size ↗

Chapter C / Analyse

Find. Examine. Verify.

LOCAL AI → REVIEW → DATASET

04 / 06 OPERATING PROCEDURE

Run AI and configure positioning.

BioSync keeps its native playback separate from the local AI process. Set a defensible survey position, then analyse supported imagery.

If the models are installed, open Tools → Settings... to view Workstation / analysis / positioning. Select Use BioSync engine if needed.
Connect or open video, then activate the AI OFF / AI ON toolbar control. Candidate detections may appear in the overlay and detection panel as analysis completes.
Use Survey → Set survey position... for a documented WGS 84 survey fix, or configure a surface NMEA input where appropriate.

Positioning: A vessel/surface GPS fix does not locate an organism underwater. Do not label a point as an organism position without suitable subsea evidence.

BioSync 2.0.1 screenshot: 04 / Analysis and positioning settings — Enable local inference and configure the positioning inputs for the survey.Open full size ↗

05 / 06 OPERATING PROCEDURE

Review the evidence, not just the box.

Treat each AI result as a candidate observation. The review workflow preserves the difference between model output and scientific identification.

Open the lower Catalogue panel and choose Review for a saved observation.
Inspect the image crop, detector score, available reference examples and any associated positioning or prediction metadata.
Enter a defensible Reviewed identity, your Reviewer name, then choose Confirm identity, Reject detection or Return to unverified. Click Save review.

Interpretation rule: A detector score is not taxonomic confidence. Unknown or ambiguous organisms can and should remain unverified.

BioSync 2.0.1 screenshot: 05 / Observation review — Inspect the stored observation and record an explicit, attributable review decision.Open full size ↗

06 / 06 OPERATING PROCEDURE

Build a reviewed dataset.

BioSync provides a separate workspace for reviewed examples, missed detections and optional downstream model-evaluation workflows.

Open Tools → Species learning / reviewed data... and select Review / missed boxes.
Correct an existing proposal or use Add missed box for an organism that detection skipped. Enter a reviewer, then choose Save species, Organism only, Background, Uncertain or Reject as appropriate.
For training work, use Footage / source splits to separate dive groups and Training / evaluation to assess candidate models before activation.

Learning is not automatic: Reviewing a crop does not instantly retrain OWLv2, BioCLIP or Qwen. Training and separate testing remain controlled steps.

BioSync 2.0.1 screenshot: 06 / Species learning workbench — Review candidates, label missed boxes and organise footage for evaluation.Open full size ↗

Final step / Deliverables

Take the observations
back to GIS.

Open the Catalogue panel and use CSV... or GeoJSON... to export locally stored observations for review, reporting or downstream analysis. Observations without suitable underwater positioning are exported without an invented location.

CSV / OBSERVATIONSGEOJSON / GISLOCAL SQLITE

Know the limits

Experimental science requires review.

BioSync 2.0.1 is an early community trial. AI boxes, species suggestions and temporary track IDs are not validated organism counts or confirmed taxonomic records.

Test your real camera and ROV feeds before a deployment. Keep source video, record human review decisions and check mapping accuracy before using exported data.