A computer vision workflow, from collection to evaluation

Start with the data you actually have

Import images, videos or an existing dataset; use a camera, a screen or, on desktop, a network camera; or play a recording through capture rules. Timelapse, motion and Smart selection help collect useful stills. Inspect the staged gallery before accepting media, then use the source and capture reason to find it again.

Prepare and structure the dataset

Crop, resize, rotate, adjust colour or redact image regions with blur, pixelation or a solid fill. Search visually by text where supported, filter media, reuse label schemas and organise several datasets in one project. Project conversion provides a preview of the changes before it modifies annotations.

Use AI with a human decision at the end

Use interactive segmentation, text or visual prompts, pose assistance, curated COCO models or a compatible custom ONNX model. Batch auto-annotation writes pending drafts for review. Imported-video predictions can be precomputed for playback and explicitly accepted; a preview is not ground truth or an automatic object-identity tracker.

Make the dataset reproducible

Run quality checks, review labels, apply image train/validation/test splits and save a dataset version. Keep source history with the media. Use project archives for portability and backups for recovery; a version on the same device is not protection from losing that device.

Measure models against reviewed test data

Evaluate an available prediction engine on annotated test images and inspect task-specific and per-label metrics. Results record their evaluation scope so changed images, annotations, schemas or splits are not silently treated as the same benchmark. Investigate weak labels, collect more useful examples and repeat the manual improvement cycle.

Export, or connect your own ML runner

Standard dataset exports work with your training tools. The experimental Pipelines integration adds a separate runner with a versioned NVIDIA TAO RT-DETR recipe for axis-aligned image detection: send a frozen dataset, request fine-tuning, inspect status and artifacts, then run inference and preview boxes. This needs a prepared runner and suitable Docker/GPU environment; it is not a hosted training service or a one-click browser trainer.

Choose it with the operating boundaries in view

• Local annotation and dataset storage stay on your device. Optional ChatGPT or Claude annotation sends the attached image/frame, labels and annotation context to the selected provider (OpenAI or Anthropic). Optional ML jobs transfer the selected data to the runner you configure; a remote runner receives that data.
• The hosted web app cannot connect to the ML runner through the current transport policy. Desktop is supported; the local browser development path is narrowly restricted.
• Network cameras need the desktop app; watched folders are not implemented. Screen capture, codecs and AI engines vary by platform.
• Runner predictions are previewed inside Pipelines. Registering the trained model in Models and accepting those predictions into Pending AI Review are not connected yet.
• There is no shared multi-user labelling queue, hosted deployment endpoint or automatic retraining loop. Store builds may trail the current source and web build.