Train and export¶
Current release 1.9.0. This is the order of work on the engineering PC, from an empty project to an exported model.
- Install and license. Install Model Bench and confirm Help ▸ Licence… shows an active licence.
-
Create a project and its classes. Each project has its own classes, images, review state and model versions. A project can hold more than one folder, so image sets stay separate. Classes can be renamed or merged later.

3. Add images. Upload Images or Upload Folder. Import COCO, YOLO or Pascal VOC if the set is already labelled. 3D Images Point Cloud Upload is for ToF captures (Pro). Use the folder filter to work on one image set at a time.
4. Label and mark reviewed. Bounding box (B), polygon (P) or rotated box (O). Only reviewed annotations train the model. Press ? for the shortcut list.
5. Train. You need at least 50 reviewed images. The app says what is missing. Training uses an NVIDIA GPU when the PC has one, otherwise the CPU. The first line of the log says which. Stop Training finishes the current epoch and saves a checkpoint. Options chosen here (package corners, Hailo-safe) shape the model and cannot be added afterwards.
6. Auto-annotate, then review. Suggestions stay unreviewed until a person checks them. Correct them, mark them reviewed, and train again. A new run warm-starts from the previous weights when the classes match.
7. Export. ONNX is Standard. Hailo and NVIDIA GPU are Pro. The file carries class names and pre-processing. The export checks the converted model against the original and fails rather than ship a model that differs. Export on the dashboard writes the labelled dataset itself (COCO, YOLO, Pascal VOC or a full ZIP), for example to share it or to train elsewhere.
8. Bring production frames back. Visual Detector can save captures.zip. Import ctrlX Captures adds each frame as unreviewed annotations underctrlx/<device>/cam<N>. Importing the same zip twice adds nothing. Correct, review, retrain, export.
Reference: editions, export targets, release notes.