CSE Doctoral Student Seminar: Golnoosh Dehghanpoor and Huayi Zeng

Apr 20, 2018
12:30 p.m.
2 p.m.
Lopata Hall, Room 101

"Tensor-Based Unsupervised Feature Learning for Satellite Imagery"

Golnoosh Dehghanpoor
Adviser: Brendan Juba

Multispectral image classification has many applications including in environmental and agricultural monitoring . To assess the changes brought about on the island of Nias near Indonesia by the 2004 tsunami and 2005 earthquake, we wish to perform change detection on the patterns of land use. Because of the scale of the problem both in time and in space, unsupervised classification needs to be performed on an appropriate feature space which takes into account different texture patterns belonging to different land covers. Unsupervised feature learning approaches that have been proposed in the literature are based on dictionary learning and auto-encoders. These approaches are iterative and converge to a local minimum. In this work, we propose a novel tensor factorization based approach for this task which obtains a unique solution. The algorithm takes the image as input and by factorizing a tensor made of patches of different sizes, outputs features for different pixels.

"Neural Procedural Reconstruction for Residential Buildings"

Huayi Zeng
Adviser: Tao Ju

Procedural modeling (PM) has revolutionized the practice of urban planning, architecture, and entertainment. PM procedurally applies shape transformation rules in a shape grammar to synthesize realistic 3D models, which have CAD quality geometries with procedural structures. Discovering such procedural structure and reconstructing CAD quality geometry from raw sensor data, such as images or 3D point-clouds, is a similar but completely different problem which we call procedural reconstruction (PR). A successful PR system could turn city-scale LiDAR scans into high-quality 3D city models with procedural structures, opening doors for novel applications. Unfortunately, most existing PR algorithms start from low-level geometry analysis in a bottom up process requiring dense and near complete 3D points. In this study, we propose a novel approach, dubbed Neural Procedural Reconstruction (NPR), which trains deep neural networks (DNNs) to procedurally apply shape grammar rules and reconstruct CAD-quality geometry models from 3D points. DNNs detect primitive structures via global analysis of entire buildings, making the reconstruction possible even from incomplete and sparse 3D data. We demonstrated the framework for a shape grammar of residential buildings in England, where LiDAR point-clouds are publicly accessible. Qualitative and quantitative evaluations over hundreds of houses demonstrate that our approach makes significant improvements over the state-of-the-art.