The workflow uses open-vocabulary object detection to generate bounding boxes from free-text prompts, which are then exported as YOLO labels and used to train a detector.

Typical workflow in the notebook:

Start with an unlabeled or weakly labeled image dataset Generate bounding boxes using prompts (for example "cat's head" or "dent in car bumper") Filter positives and rebalance the dataset Export labels in YOLO format Train and evaluate a YOLO model

In the example notebook I use a cats vs dogs dataset with only image-level labels. Using the prompt "cat's and dog's head", the pipeline auto-generates head bounding boxes and trains a small YOLO model.

The repository mainly contains the Colab notebook plus example scripts for running the detection and exporting YOLO labels.

Curious to know what people think of this approach.