Speaker: Hana Lebeta Goshu

Abstract: Three dimensional (3D) computer vision provides essential spatial information for autonomous navigation, robotics, augmented reality, environmental monitoring, and infrastructure assessment by reconstructing and interpreting physical environments from two dimensional images. Recent advances in neural scene representation have progressed from the implicit volumetric functions of Neural Radiance Fields (NeRF) to the explicit anisotropic primitives of 3D Gaussian Splatting (3DGS), enabling high quality reconstruction and real-time rendering. However, limitations at the detection, initialization, optimization, densification, and composition stages reduce reconstruction accuracy, stability, and computational efficiency. Accordingly, this seminar presents artificial intelligence (AI) methods developed to address limitations across NeRF and 3DGS. First, an attention-driven NeRF framework models directional scene structure through axis specific attention and volumetric feature fusion, improving 3D object detection across two indoor scene benchmarks and multiple neural network backbones. Next, parameter averaging and multiresolution supervision stabilize joint optimization for egocentric dynamic reconstruction, improving reconstruction quality without additional inference cost. Moreover, stereo depth is converted into explicit geometry through point injection, within view quality filtering, and virtual view supervision while preserving or improving surface reconstruction accuracy. Finally, image structure guides adaptive densification, while full covariance anisotropic transmittance filtering reduces aliasing under changes in rendering scale. Overall, the methods form a coherent progression from implicit detection to explicit reconstruction and motivate a unified framework for reconstruction, recognition, uncertainty estimation, and computational efficiency. Ultimately, the research supports reliable AI systems for accurate 3D perception under sparse observations, dynamic environments, and real time rendering requirements.

Bio: Hana Lebeta Goshu’s research lies at the intersection of artificial intelligence (AI), machine learning, computer vision, and computer graphics. Her expertise spans 3D computer vision, neural scene representations, optimization, physics-informed AI, generative AI, deep representation learning, and uncertainty quantification. Her research develops reliable and computationally efficient methods for 3D scene reconstruction and understanding, with particular emphasis on Neural Radiance Fields, 3D Gaussian Splatting, sparse observations, and 3D object detection. She has developed and publicly released multiple research projects and authored eleven peer-reviewed publications in venues including ACM Multimedia, Neurocomputing, Applied Soft Computing, and Engineering Applications of Artificial Intelligence. She was one of 11 recipients worldwide of the 2025 IEEE Computational Intelligence Society Graduate Student Research Grant. Her academic experience includes seven terms of undergraduate and postgraduate teaching, project supervision, and mentorship of more than 15 students. She has earned multiple awards recognizing academic excellence, research, leadership, and service while maintaining first-class academic standing throughout her university studies.

https://us06web.zoom.us/j/84636366540
MEETING ID: 84636366540
PASSCODE: 25862