NeRF (neural radiance fields) is a method for synthesizing novel views of complex scenes by optimizing an underlying continuous volumetric scene function using a set of input views.
This function is represented by a simple MLP (multi-layer perceptron network) which holds the scene geometry in its learned weights.
We intend to simplify NeRF even more and use a generative approach to generate novel scenes.
Though a lot of work has been done for generative models in 2D images, few have researched the 3D task.
Our network will train multiple scenes together to create a latent space, where each scene is represented by a latent vector (or image). Using the latent space and a decoder, we will generate new scenes or condition on existing scenes using very few images.
I finished my Master's degree in Computer Science at the university of Haifa in 2020, under the supervision of prof. Hagit Hel-Or.
Forgery detection in 3D-sensor Images
The field of Image Forensic, and with it the notion of image forgery and its
detection, is widely studied in 2D images and videos. With the increase in avail-
ability and use of cameras with depth sensors, it has become necessary to consider forgery detection in depth-images as well. In this research, we present an introductory study of forgery detection in depth-images. Specifically, we show that noise statistics in depth-images can be exploited for camera source identification, image forgery detection. We further show that scene illumination can be used to detect forgery. Finally, we show that inherent characteristics of the camera mechanics can be exploited to determine image forgery from sensor-based shadows.
Fun github projects
Isn't it annoying when a GIF is jumpy? Meaning, the first and last frame are not the same and the video jumps from end to beginning.
In this project, we fix it by learning the scene using NeRF technique and correct the camera path around the object to create a perfect loop. Now you can enjoy a corrected 360 view around your object.
By looking at two painting, can you tell if they belong to the same artist?
Well, our network can. A trained Siamese CNN can output a score for painting and identify which paintings were painted by the same artist.
This challenge was introduced by Kaggle. The suggested solutions used a classifier with labels which classifies images to a set of determined artists. Our solution doesn't need labels in the inference stage and can score paintings by new artists not seen in training.