Projects Last Projects
Precise Completion of Deformable 3d Shapes Using SIREN
Supervised by Oshri Halimi
With VR headsetâ€™s popularity growing and its application for VR chats raising interest, the need for a fast, cheap and quality way to transform the userâ€™s shape and movements in real time to the VR space arises.
In this paper I will propose my idea for a network that can reconstruct a partial scan from a single angle which came from a depth sensor to a full precise 3D shape.
I will discuss the representation of a 3D shape as an SDF and its possible advantages for shape completion, along with SIREN network that uses an easy to calculate Loss function and shows how the use of periodic activation functions such as Sine can improve both the speed and the quality of fitting a point cloud to an SDF.
Finally, I will show how I used SIREN network and SDF representation to implement a new network for the use of precise completion of deformable 3d shapes.
Please, see project report.
Please, see final presentation.