VisionEngine demo: descriptive text alternative

This recording has no audio. It follows an example object-detection project,
"Site safety", from photographs of workers wearing hardhats to a trained model.
It is a retained recording: interface details, account badges, credit balances
and estimates shown are historical, not current offers. The example results are
not guaranteed performance on other data.

Upload
The demonstrator creates a project, selects "Detect objects" and names it
"Site safety". The Build page then shows 14 images uploaded. The photographs
contain workers and hardhats in different positions and settings.

Baseline
Under "Set up your baseline", the demonstrator selects Text prompt and enters
"hardhat" as the object to detect. A Visual example option is also available.
The baseline runs over the images. When it finishes, the gallery shows suggested
boxes around hardhats and separates saved labels from predictions.

Prioritise
The gallery is sorted with the most uncertain images first. An uncertainty
threshold slider separates suggestions into an auto-accept band and a review
queue. Moving the threshold changes the counts and which images appear in each
band; it does not mean every suggestion has been manually checked. Image cards
carry priority scores and label/prediction counts. The demo returns the
threshold to about 0.29 before reviewing images.

Review
Opening an image brings up a large photograph with boxes around the hardhats
and a side panel listing suggestions. Accept all, Reject all and individual
accept/reject controls are available. On-screen instructions explain that
accepted boxes become training labels and a missing detection can be added by
drawing a box. The demonstrator accepts suggestions and moves between images;
a later photograph shows an accepted box around a white hardhat. Returning to
the gallery shows 6 of 14 images reviewed, with labels and remaining predictions
counted separately.

Train
The demonstrator opens "Train a model". The dialogue identifies 9 labelled
images of hardhats. Its advanced settings show training epochs, a held-out
validation proportion and augmentation. In the recording the settings reach
10 epochs, 30% held out for validation and augmentation enabled. The dialogue
shows a time and credit estimate, then training progresses to "Trained v1".

Measure
The completion dialogue displays mAP50 0.828, precision 0.660 and recall 1.000.
"View in Models" opens the trained-model list. Version v1 is marked Ready and
Live, with the same scores. A warning explains that results from a
small validation set may be unreliable and a single prediction can swing them;
labelling more images can support a more stable evaluation. A labelset-history
table records the training version and image/label counts. Evaluate and Try on
new images are visible next steps; this recording ends on the Models page.