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.
- 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.
- 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.