Semantic Understanding of Indoor Scenes Using Deep-Learning

Abstract

Nowadays, indoor scene understanding, and segmenting are a prerequisite for several commercial and educational applications in the furniture industry, interior home design and remodel. That is why it considered as a key computer vision area of research. In this thesis, a deep learning approach and its algorithms are employed to introduce a fully automatic system to understand and segment a variety of living rooms and bedrooms styles. First, a classification Convolutional Neural Network “CNN” model architecture was implemented to distinguish between main home furniture which are chairs, beds, and tables. After that, further step is taken to achieve the target of semantic indoor scenes understanding. This step is using a deep learning segmentation algorithm which is Fully Convolutional Network “FCN”. The algorithm consists of down-sampling phase, up-sampling phase, and fusing phase. Through the down-sampling phase a customized architecture and a pretrained architecture which is VGG-16 were implemented. The segmentation FCN models were built and evaluated on Keras and TensorFlow frameworks. Regarding the evaluation, the CNN and FCN models were subjective, and objective evaluated by calculating the training and validation loss and accuracy and compare the results to each other to select the most promising model for the target classification and semantic segmentation task. The CNN model was proven its capability of classification for the selected classes with 79.96% training accuracy and the FCN model with pre-trained VGG-16 was proven its capability of semantic segmentation for main living rooms and bedrooms furniture with 94.18% training accuracy which were satisfying results.