Skip to content

Workflow and deployment

A project goes from raw images to a model running on ctrlX in seven steps, and production frames flow back to improve it.

  1. Create a project and its classes in the dashboard (e.g. scratch, parcel, missing_screw). Classes can be renamed and merged later.
  2. Add images: Upload Images or Upload Folder; Import COCO for labelled datasets; Import ctrlX Captures; 3D Images Point Cloud Upload for ToF captures (Pro).
  3. Annotate with bounding box (B), rotated box or polygon (P). Only annotations marked Reviewed train the model.
  4. Train Model on the reviewed annotations: choose model size, epochs and train/validation split, and follow the live log. Options: package corners, Hailo-safe (Pro), object outlines (desktop only, not exportable yet).
  5. Auto-Annotate the remaining images. Suggestions arrive unreviewed; correct them, mark them reviewed, and train again.
  6. Export ONNX (for ctrlX) to the chosen target. Every export runs the converted model against the original and fails rather than ship a model that differs.
  7. Import ctrlX Captures: the Visual Detector saves live frames with its detections as captures.zip; importing it adds each frame with unreviewed annotations, records device, camera, exposure and frame number, and files it under ctrlx/<device>/cam<N>. Importing the same zip twice adds nothing.

Screenshots to come

  • S8 — Project dashboard with projects and the toolbar.
  • S9 — Annotation canvas: rotated box or polygon, one annotation selected, class panel visible.
  • S10 — Train Model dialog with the live log during training.
  • S11 — Auto-annotation: machine suggestions beside reviewed annotations.
  • S13 — Dashboard after Import ctrlX Captures.