Project Indigo
Experimental iOS computational photography app and research platform for ML-based multi-frame fusion, color correction, super-resolution, and related imaging applications, serving as a testbed for on-device model deployment.
Ilya
Chugunov
About: I'm a Research Scientist based in Seattle, working with the Adobe Nextcam team on computational photography. I received my Ph.D. from Princeton University, where I was part of the Princeton Computational Imaging Lab advised by Professor Felix Heide, and was supported by the NSF Graduate Research Fellowship. I earned my bachelor's degree in electrical engineering and computer science from UC Berkeley.
Contact: cout
<< "contact" << "@" << "ilyac.info"
Research: I'm interested in computational photography, 3D reconstruction, and inverse problems that look at the whole imaging pipeline, from signal collection to scene reconstruction. MRIs, modulated light sources, or mobile phones; I love working with real devices and real data.
Over the course of my research I've developed several open-source apps for data collection:
- Pani : Android app for recording of multi-camera Bayer RAWs and metadata.
- SoaP : iOS app for recording Bayer RAWs, depth maps, and metadata.
- HNDR : iOS app for recording video, depth maps, and pose estimates (from ARKit world tracking).
"If you try and take a cat apart to see how it works, the first thing you have on your hands is a non-working cat."
- Douglas Adams
Hist2Style: Histogram-Guided Stylization
Lightweight network which transfers the color look of a reference photo in real time, conditioned on a histogram embedding of the target.
Seeing Through Fibers: Unsupervised Image Reconstruction in Fiber Bundle Imaging Systems
Test-time training method which reconstructs fiber bundle endoscope images by jointly estimating motion and scene, with no calibration or paired data.
Neural Atlas Graphs for Dynamic Scene Decomposition and Editing
2.5D scene representation which models video as moving planes in 3D with learned atlases, supporting appearance edits and depth re-ordering.
Neural Field Representations of Mobile Computational Photography
Neural Light Spheres for Implicit Image Stitching and View Synthesis
Spherical light field model which stitches panoramas while handling parallax, view-dependent shading, and scene motion at real-time 1080p.
Split-Aperture 2-in-1 Computational Cameras
Camera design which codes half the aperture with a diffractive element, capturing coded and conventional images at once for HDR, spectral, or depth imaging.
Neural Spline Fields for Burst Image Fusion and Layer Separation
Motion representation which models burst pixel trajectories as splines, enabling image fusion from photometric loss alone and separation of reflections and shadows.
Shakes on a Plane: Unsupervised Depth Estimation from Unstabilized Photography
Unsupervised method which recovers depth and camera motion from the hand tremor parallax in a two-second RAW burst, without LiDAR or pose estimates.
GenSDF: Two-Stage Learning of Generalizable Signed Distance Functions
Two-stage meta-learning approach which learns generic shape priors, allowing signed distance function reconstruction of over a hundred unseen object classes.
Centimeter-Wave Free-Space Time-of-Flight Imaging
The first free-space GHz time-of-flight imager, built with electro-optic modulators and paired with a network for unwrapping ambiguous phase.
The Implicit Values of A Good Hand Shake: Handheld Multi-Frame Neural Depth Refinement
Refinement method which combines phone RGB frames, ARKit poses, and coarse depth to produce a high-fidelity depth map from a single snapshot.
Mask-ToF: Learning Microlens Masks for Flying Pixel Correction in Time-of-Flight Imaging
Jointly learned microlens mask and decoding network which suppresses flying-pixel artifacts in time-of-flight depth while remaining light efficient.
Self-Contained Jupyter Notebook Labs Promote Scalable Signal Processing Education
Education study arguing that self-contained notebook labs match in-person signal processing sections while reducing course staff overhead.
Multiscale Low-Rank Matrix Decomposition for Reconstruction of Accelerated Cardiac CEST MRI
Reconstruction method which exploits Z-spectrum sparsity for 4-fold accelerated cardiac CEST scans with accurate Lorentzian line-fit analysis.
Duodepth: Static Gesture Recognition Via Dual Depth Sensors
Gesture recognition approach which fuses dual depth sensor point clouds via a 3D spatial transformer network, outperforming explicit ICP registration.