Hide/Show Apps

3d indoor scene segmentation using consensus clustering

Küçükdemir, Furkan Mevlüt
In this study, we propose an indoor scene segmentation method which utilizes consensus clustering. Unlike most of the recently offered methods in literature, our approach does not rely on deep learning techniques. Therefore, it does not require a large dataset for training and spending a lot of time to learn a valid model is not necessary. In the first step of our algorithm, we construct uniform and cotangent Laplace operators. Then, we compute differential coordinates using them and global point signatures using the eigenbasis of cotangent Laplace operator. In the next step, we use these coordinates and global point signatures as features and run k-medoids multiple times to create an ensemble. With the help of Partial Evidence Accumulation Clustering method, which is a consensus clustering approach, we obtain the final segmentation. Optionally, we offer an interactive segmentation mechanism to our users, in case any adjustment on the final segmentation is needed. The key idea of our approach is to use a specialized function to compute distance between feature points, which takes the scene geometry into account by using surface properties such as normals. At the end of the thesis, we also present both qualitative and quantitative evaluation of our method and show that it outperforms some of the existing techniques, quantitatively.