I am an Applied Research Scientist at Geomagical Labs. I received my PhD from Cornell Tech, where I was advised by Ramin Zabih. I was also a Student Researcher at Google Research. I am interested in computer vision; I have worked on a variety of vision problems and am interested in self-supervision, depth estimation and 3D. Prior to graduate school, I worked in various places in industry, most recently at Google on the VR Jump camera (I am the engineer who is “weirdly” riding a unicycle in this Wired article). I earned my BS in math (with honors and departmental distinction) from Harvey Mudd College in 2010. I also author crosswords.
Publications
DreamWalk: Style Space Exploration using Diffusion Guidance
ECCV 2024 AI4VA workshop, project page M. Shu, C.Herrmann, R.S. Bowen, F. Cole and Ramin Zabih
We show that by varying guidance scales over time and space, we can achieve more fine-grained control of image style and content when generating with a standard diffusion model.
Dimensions of Motion: Monocular Prediction through Flow Subspaces
3DV 2022 (oral), project page R.S. Bowen, R. Tucker, R. Zabih, and Noah Snavely
We developed a new method for self-supervised monocular disparity estimation from a video collection that skips the traditional posing step, by casting depth estimation as a subspace-estimation problem for optical flow.
Deep Survival Analysis with Longitudinal X-rays for COVID-19
ICCV 2021 M. Shu, R.S. Bowen, C. Herrmann, G. Qi, M. Santacatterina, and Ramin Zabih
Using a imaging dataset of early-pandemic COVID-19 patients, we use time-to-event analysis to do automated prognosis from chest x-ray and other data, even in the presence of right censoring.
OCONet: Image Extrapolation by Object Completion
CVPR 2021 R.S. Bowen, H. Chang, C. Herrmann, P. Teterwak, C. Liu, and Ramin Zabih
The image extrapolation problem is particularly challenging when the border cuts through a semantic object. We show substantial improvements by using a two-armed pipeline, with one arm specifically trained to complete objects.
Finding Maximally Informative Patches in Images
NeurIPS 2021 Workshop Paper H. Zhong, G. Balakrishnan, R.S. Bowen, R. Zabih and William T Freeman
We address the problem of finding maximally-informative patches, i.e., a few patches from a given image from which reconstructing the rest of the image is easy.
Learning to Autofocus
CVPR 2020, project page C. Herrmann, R.S. Bowen, N. Wadhwa, R. Garg, Q. He, J.T. Barron, and Ramin Zabih
Autofocus for mobile-device cameras typically relies on classical algorithms based on, e.g., local contrast metrics. We produce a dataset and demonstrate how to train a network to improve focal-distance estimation.
Channel Selection using Gumbel Softmax
ECCV 2020 C. Herrmann, R.S. Bowen, and Ramin Zabih
The Gumbel distribution allows us to reparameterize a discrete distribution; we use the so-called straight-through trick for network pruning, either as a model compression step or adaptively at runtime based on the input.
Robust Image Stitching with Multiple Registrations
ECCV 2018 C. Herrmann, C. Wang, R.S. Bowen, E. Keyder, M. Krainin, C. Liu, and Ramin Zabih
Object-centered Image Stitching
ECCV 2018 C. Herrmann, C. Wang, R.S. Bowen, E. Keyder, and Ramin Zabih
These two papers address panoramic image stitching in an object-centric way, updating the classic stitching pipeline to better handle moving objects.
Teaching
I was involved in the creation of CS 5112, an algorithms class for masters students at Cornell Tech, taught by my advisor. I have also been a teaching assistant for CS 2112, an honors data structures course for Cornell undergrads taught by Dexter Kozen. As an undergrad, I was a frequent teaching assistant for CS 81, the introductory theory course.