Tailoring the AI for Robotics: Fine-tuning Predefined Deep Convolutional Neural Network Model for a Narrower Class of Objects

Object recognition has been one of the greatest challenges for robotic tasks particularly in indoor environments. In order to perform this task, so called hand-crafted features requiring high computational effort which rely on expert knowledge are used until achieving better results by learned features using convolutional neural networks. In this study, we have given a powerful object classification model with 2.72% top-1 error rate which is achieved by fine-tuning a predefined model for 10 classes. We built our model on top of VGG16 architecture (trained on a larger dataset which consists of 1000 classes) and froze the layers except the last classification layer which we trained for 10 classes. The training data consists of 10000 images (1000 per class) and there are 4000 images for validation (400 per class). The object classes in our dataset are book, bottle, bowl, cup, eyeglass, keyboard, laptop, monitor, teapot and vase which may exist on the desktop in an indoor environment.