Visual Context Augmentation

  • @dvornikModelingVisualContext2018
  • learns to place object instances at an image location depending on the surrounding context.
  • A neural network is trained for this purpose.
  • The training data is pre- pared to generate a context image with the masked- out object inside it.
  • From an image, 200 context sub-images are generated surrounding the blacked- out bounding box. The neural network learns to predict the category (object or background) in masked pixels.
  • The object instances are placed inside the selected boxes to generate a new train- ing image