3D Scene Recognition From a Single Image

Human eyes capture the world around us and effortlessly derive an impression of scene depth from a single image. However, developing an artificial system that can identify the impression of the 3D scene with the same performance and robustness as humans, still is a challenge for researchers from such fields as physiology, computer science, and artificial intelligence. The 3D scene recognition from a single image is an important problem for many applications of computer vision such as autonomous vehicle control, scene understanding, and 3D TV. The contributions of the thesis are explored in three different ways. First, the segmentation-based feature extraction method is introduced to classify the relatively clear geometry structure images, in which the image features are extracted by exploiting predefined templates, each associated with an individual classifier. Each of the individual classifiers learns a discriminative model and their outcome are fused together using sum-rule for recognizing the 3D scene geometry of an input image. It achieves 86.25% recognition accuracy on ‘stage dataset 1’, which is higher than the state-of-the-art methods. In the second contribution, a new method of 3D scene recognition-based on the fusion of deep convolutional neural network (CNN) features and texture gradient features is presented. Meanwhile, as the 3D scene geometry dataset is not publically given, thus, a medium scale, ‘stage dataset 2’, is introduced. Experimental results exhibit that the proposed method reaches 86.29% recognition accuracy, which achieves higher accuracy and faster than the baseline methods. Finally, in the third contribution, the handcrafted features are integrated with multi layer features at different intermediate blocks of CNN, and each block is connected with an individual classifier and then scores of these classifiers are combined while using sum and product-rule to recognize the scene geometry type. The introduced approach is validated on two benchmark datasets and it achieves 95.17% and 97.68% recognition accuracy on ‘stage 2 dataset’ and ‘15-scene’, which is superior to the state of-the-art methods. Keywords: CNN, Ensemble of classifiers, Handcrafted feature, Multi-layer features, Predefined templates, Stages, 3D scene recognition