Coping with Overfitting Problems of Image Caption Models for Service Robotics Applications

Abstract

Image captioning is a task for generating the sentence to properly describe the image. This is a difficult task because multiple concepts should be captured by the model from the given images, such as object recognition, human detection, action recognition, etc. Therefore, if using the traditional data-driven approach to solve this problem, there will be an upper limit of the performance due to the random selection for training images. Most of the case, overfitting is a major problem. Therefore, we propose several indicators and observe some insights about overfitting in image captioning tasks. Finally, we propose a dynamically increasing threshold method to observe the stability and reliability of the image captioning model.