Automatic labeling of infrared images

Background & motivation

  • The majority of object detection methods for camera images are based on data-driven methods such as convolutional neural networks that requires large amounts of labeled data.
  • Labeling images manually is a tedious an non-scalable task, and (semi)automated labeling processes are desirable.
  • The vast majority of available networks are based on regular RGB images. An alternative camera type with useful image information are infrared (IR) cameras, which sense the temperature of objects.

Problem description

  • The research vessel milliAmpere has a sensor rig that contains both infrared and RGB cameras. The cameras are stacked vertically, which means that the cameras have a large overlap in the field of view. This means that objects detected in the RGB frame can be transferred into the IR camera image.
  • This way, the relatively weaker detection methods used in IR images can be trained using this automatically labeled data. 

Work proposal

  • Implement the detection transformation from RGB to IR images.
  • Train the IR detector using the automatically labeled images.
  • Compare the performance of the detectors on a test set.
  • Write report.

Figure 1: An image of the sensor rig on milliAmpere.