Motivation

• Recent research has shown there is still plenty of room to grow model performance through augmenting our training data. Roboflow has written extensively about data augmentation and has highlighted some of the recent advances that have made new models like YOLOv4 and YOLOv5 state of the art.

Objectives

• The mosaic augmentation was invented by Glenn Jocher earlier this year and was first released in YOLO v4. 

Related Works

Idea

• It works by taking four source images and combining them together into one. 

• This does a few things:

  1. It simulates four random crops (while maintaining the relative scale of your objects compared to the image) which can help your model perform better in cases of occlusion and translation.

  2. It combines classes that may not be seen together in your training set (for example, if you have pictures of apples and pictures of oranges, but no pictures of apples with oranges in the same photo, mosaic will simulate that).

  3. It varies the number of objects in your images (for example, if all of your images only contain one bounding box, the output of the mosaic will have between zero and four).

When to Use

• Aerial Imagery.

• Real World Objects.

• Low Object Distribution

When Not to Use

• Written Documents.

• Large, Prominent Objects.

• Fixed Location Objects.

Dataset

Metrics

Experimental Results

Ablations