Gene Chou

I am a second year CS PhD student at Cornell University, advised by Noah Snavely and Bharath Hariharan. Previously, I received my bachelor's in computer science and applied math from Princeton University and worked with Felix Heide. My research focuses on 3D reconstruction and generation. I am supported by the NSF graduate fellowship.

Contact: gene@cs.cornell.edu

We propose the task of generating videos from sparse, unposed internet photos, and design a self-supervised method that takes advantage of the consistency of videos and variability of multiview internet photos to train a 3D-aware video model without any 3D annotations such as camera parameters.

MegaScenes is a scene-level dataset containing 100K SfM reconstructions and 2M registered images, collected from Wikimedia Commons. We validate its effectiveness in training large-scale, generalizable models on the task of single image novel view synthesis.

Builds on YOLOR to jointly train vision (e.g. object detection, instance and semantic segmentation) and vision-language (e.g. image captioning) tasks. Fast and lightweight while achieving competitive performance.

Recovers images in broadband using a single flat metasurface optic. Compensates for residual aberrations with probabilistic deconvolution implemented using a conditional diffusion model.

Performs diffusion on the latent space of neural SDFs while providing geometric guidance. Generates diverse meshes conditioned on partial point clouds, 2D images, and real-scanned, noisy point clouds.

Combines a semi-supervised approach with a self-supervised loss to reconstruct neural SDFs from raw input point clouds of over a hundred unseen object classes.

Website template borrowed from Jon Barron.