The pre-trained TensorFlow Object Detection models certainly work well for some problems. But sometimes, you may need to use your own annotated dataset (with bounding boxes around objects or parts of objects that are of particular interest to you) and retrain an existing model so it can more accurately detect a different set of object classes.
We’ll use the same Oxford-IIIT Pets dataset, as documented in the TensorFlow Object Detection API site, to retrain two existing models on your local machine, instead of using Google Cloud covered in the documentation. We’ll also add an explanation for each step when needed. The following is the step-by-step guide on how to retrain a TensorFlow object detection model using the Oxford Pets dataset:
- In a Terminal window, preferably on our GPU-powered Ubuntu to make the retraining...