Application Of Subspace Learning To Kinect Based Scene Classification

With the rapid development of digital image processing and Internet technology, thescene classification problem has obtained intense attention, and become one of the hot field inimage processing technology research. Despite the continuous development of research oncomputer vision and the robot in recent years, Scene classification remains to be a complexproblem which fails to be well solved, which could be due to the complexity of the actualenvironment, illumination in the scene and unpredictability of scene entities.In this work, the main contributions are listed as follows:1. In this work we discuss the effect on the scene classification accuracy when the depthinformation of image is taken into consideration. Before Microsoft Launched the Kinect, mostof the scene classification researches are based on2D color images. In this work, ourexperiments are based on the NYU data set, which combines RGB color information anddepth information, to gain an insight of the effect that the depth information has on the sceneclassification recognition rate.2. We discuss the influence of different feature representation strategies on therecognition accuracy based on SIFT. The feature representation strategies we discuss includeScSPM, Locality-constrained Linear Coding (LLC), Efficient Match Kernels (EMK), etc.3. We use subspace learning algorithms to decrease the computation cost under ScSPM,LLC and EMK high dimensional feature representation strategies, and propose a newsubspace learning algorithm named Rank Preserving Discriminant Analysis (RPDA) based onPatch Alignment Framework. Experiments are conducted to find out the effectiveness ofseveral different subspace learning algorithms like RPDA, PCA, LDA and MFA with differentfeature representations in scene classification.