Reinforcement Learning for Improving Object Detection

Authors: Siddharth Nayak, Balaraman Ravindran

We use reinforcement learning(RL) to make digital transformations in images to extract the maximum object detection performance from a pre-trained network. We work with digital distortions like brightness, colour and contrast of the image. [PDF], [slides], [ECCV presentation]

In the above figure, the left column has images from the PascalVOC dataset and the right column has images obtained after the trained RL agent applies the digital transformation (brightness in this case). In the first example, only one person is detected in the input image whereas two persons are detected in the agent-obtained image. Although, the original image looks pleasing to a human eye and the agent-obtained image is quite bright, the agent-obtained image performs better in detection. Similarly, in the second example, the detector misses the group of people in the input image. It detects that group after the agent transformation.