It seems you trained the mode, fine-tuned the mode.

I was wondering, I can do directly on TFJS, see my post below. Why should I pass by all those steps?

obs. In summary, if you prioritize accuracy, YOLOv7 is an excellent choice. However, if real-time performance on resource-constrained devices is crucial, consider using MobileNet SSD v2

https://jorgeguerrapiresphd.editorx.io/tensorflowjs-academy/post/detecting-and-highlighing-objects-using-ssd_mobilenet_v2