Semantic Image Segmentation with Deep Features

Deep convolutional neural networks (CNN) have shown significant success in many classification problems including semantic image segmentation. However training of deep networks is time consuming and requires large training datasets. A network trained in one dataset could be applied to another task or dataset through transfer learning and retraining. As an alternative to transfer learning, feature vectors that are extracted from network layers could be directly used for classification purposes. In this paper we investigate the improvement in classification performance when features extracted from generic CNN architectures are used in an image labeling algorithm that does not require training. We show that the use of 'learned' features from deep networks together with 'hand-crafted' features such as SIFT increases the labeling accuracy. Since existing pre-trained networks are used, the proposed approach could be easily applied to any dataset without any retraining. The proposed method is tested in two datasets and labeling accuracies are compared with similar existing methods.