Before starting, ensure that your datasets are converted into the COCO format (JSON file). The COCO format is widely used for object detection tasks and consists of three main components:
images: Contains the metadata of the images (ID, file name, dimensions, etc.).
annotations: Contains the annotations for each image, including the bounding box coordinates, segmentation masks, and category IDs.
categories: Contains the list of object categories (classes).
You can use custom scripts or available tools to convert datasets from various formats (Pascal VOC, YOLO, etc.) into the COCO format. Once the datasets are in COCO format, proceed to the next step.
After converting your dataset into COCO format, use the provided split.py script to split the dataset into training and testing sets. The script randomly divides the dataset into the specified proportions (e.g., 80% for training and 20% for testing).
Here’s how to use the split.py:
This script will output two new JSON files:
1.train.json: Contains the training data.
2.test.json: Contains the testing data.
(If the dataset has its own partitions, we follow their original partitions)
After splitting the dataset, you can select a suitable number of samples from the training set for k-shot learning. k-shot refers to selecting a small number (k) of examples for each category in the dataset.
To generate a k-shot dataset from the training set, follow these steps:
2.Run a custom script to extract k-shot samples for each category from the training dataset. You can set k to 1, 5 or 10 as needed for your k-shot learning task.（make sure always to be one annotation in a picture）
Once this process is complete, the kshot_train.json will contain the selected k-shot data, which can be used for training k-shot models.
Edit and run build_prototypes.sh.
bash build_prototypes.sh
Before training, update these two files: lib/categories.py, detectron2/data/datasets/build.py. Then adding your datasets name in 'datasets_name'.