IPAB Workshop - 25/07/19

Geng Lyu

Title: Microscopy level capturing and representing of curved surface.

Abstract: We present an approach to capturing and representing the details of a curved surface. The existing 3D scanning methods based on laser scanner or RGBD sensor fails to provide high resolution textures while microscopy scanning methods used in biological research are usually designed for planar surfaces. We combine motion sensor with a portable microscope to get the rough pose of the captured microscope images and then apply global pose optimization to make a microscope level 3D reconstruction of the scanned object.

Emmanuel Mbabazi

Title:  Lower Dimensional Kernels for Video Discriminators

Abstract: Traditional video GANs use discriminators with 3D kernels when estimating the probability that an input video is real or fake. The intuition is that 3D convolutions can better capture the temporal information in video, allowing for more accurate criticism of generated video. The challenge with using 3D convolutions, is that they become prohibitively memory and compute expensive as network complexity increases. In this work, we investigate how to overcome this challenge by replacing 3D kernels in TGAN discriminators with lower dimensional approximations. Our findings show that discriminators with factorized kernels provide a better learning signal during GAN training than their 3D counterparts. As a result, we propose a factorized discriminator architecture that improves performance, while being more parameter efficient.