Object detection locates target objects in images by pairing each object with a class label and a bounding box. A production workflow therefore combines image collection, human labeling, dataset versioning, and model training. section 2.3.2
A practical automated workflow separates responsibilities into stages:
- Store incoming images in a raw image dataset. section 2.3.2
- Launch a labeling application that lets users draw bounding boxes around target objects. section 2.3.2
- Save images and generated label files in a labeled dataset. section 2.3.2
- Trigger model training when new labeled data arrives, then save the model file. section 2.3.2
- Version raw data, labels, and models so provenance connects a model to the exact training data and processing code. section 2.3.2
