[Feature]: Data augmentation method by extracting objects and placing them on different backgrounds

[TomyangSydney]

Guidelines

• I agree to follow this project's Contributing Guidelines.

Description

We can try to generate some new images containing object with insufficient annotations. To generate new images, we can extract trash objects from the current background and place them on the background of other images and save it. Then we can use them to train the model.

Problem

Insufficient amount of data for several trash categories in TACO dataset.

Proposed Solution

We can build a pipeline to generate new images

1. Extract the trash from existing images from the backgrounds based on the information from the annotation.json file
2. Attempt to fill in the holes from the backgrounds and store new backgrounds
3. Mix and match the trash onto the different backgrounds and save the images

We can combine those new images with the official TACO photos to train the Yolov5 network.

Alternatives Considered

1. Obtain new images, apply manual annotations and add these images for training
2. Use regular data augmentation method such as rotating the image
3. Try other dataset