13.7.6. Wrap up

The whole path, end to end: record footage on the camera, upload the clip and sample it into frames, label the frames with Auto Label and a review pass, build a dataset version with sensible augmentations, train a small YOLO-family detector on Roboflow’s servers, and download integer-quantized TFLite weights from the OpenMV deploy target. The model then loads through the same ROMFS flow as any other.

One idea runs through all of it: train on what the camera will actually see. Capture with the camera that will run the model, augment for the variations it will meet, and train at a small resolution that fits the hardware. A model built that way performs on the camera the way it did in testing.