Dr. Supasorn Suwajanakorn
- Ph.D. in Computer Science, University of Washington
Lecturer at School of Information Science and Technology (IST), Vidyasirimedhi Institute of Science and Technology, Rayong, Thailand
- Supervising 9 Ph.D. students
We explore the problem of view synthesis which allows for the generation of new views of objects or scenes given from a set of few input images. We focus on generating high-quality view extrapolations. This is challenging, it requires comprehensively understanding the 3D scene from 2D images. In this context, occlusions and depth uncertainty are also the most pressing issues, and worsen as the degree of extrapolation increases. We’re looking for a practical and robust deep learning solution for capturing and rendering novel views of complex real-world scenes.
With the current state-of-the art technique for 3-D modeling it is not good enough many thing such as, glass, fluffy things, moving people. So, our group delve into new technique called Multiplane Images (MPIs) to tackle this problem.
Improve the quality of Time-Lapse Reconstruction from Internet Photos
We want to try using GANs and other methods to improve the quality of Time-lapse generated from internet photos instead of running a camera at the same place for a long time. We aim to solve three main problems from the existing work. Firstly, the generated Time-lapse has low resolution. Secondly, sometimes thin objects disappear due to the different perspectives from varied photos. Lastly, we want to generate photos from the period that there are no photographs taken to reduce the “skip” feeling of Time-lapse video.
Unsupervised Landmark Detection via Spatial Reasoning
Creating annotation for landmark detection requires high human labor. Hence, this work aims to create a network that can learn to detect landmarks of an object unsupervisedly. The network learns the spatial relationships between components within an object and tries to apply the acquired knowledge on the landmark detection task without human annotation.
Retrieving real video from DeepFake
DeepFake can be used in harmful ways for example: delivery misleading political speech, cyberbullying, etc. What is DeepFake? Basically it is a video generated by an “AI”(Face2Face, FaceSwap, DeepFake). In order to generate a new DeepFake video it requires an original video as a base or source video for an AI. By retrieving an original video from a DeepFake we can show that the video is being manipulated by DeepFake and help combating DeepFake videos.
Set representation learning for retrieval task
Set is a type of data that forms a collection of things. Such as set of points in 2-dimensional space, a set of points in 3-dimensional space or a point cloud, a set of objects in each image, etc. Unfortunately, applying deep learning with the set is a challenge issue because the set has a property called permutation invariant which isn't applicable with traditional neural network approaches that treat each sequence or position of vector differently (i.e., permutation equivariant). For this reason, we try to design a neural network that can represent the set correctly which can extend the ability to apply set to other several tasks like classification, retrieval, and segmentation.