Leonid Keselman

I am a PhD student at the Robotics Institute, part of the School of Computer Science at Carnegie Mellon University, where I work on 3D computer vision. My PhD advisor is Martial Hebert.

From 2011 to 2017, I worked at Intel, as part of Intel RealSense. I primarily designed computer vision algorithms for efficient hardware ASICs, including the Intel RealSense R200 and D400 RGB-D sensors. Additionally, I worked on software APIs, active illumination systems, human-computer interaction devices, and helped develop demos for trade shows, including CES 2012-2016.

I have an MS in Computer Science (AI focus) from Stanford University, where I was a research assistant for Silvio Savarese and a teaching assistant for Fei-Fei Li (CS131 & CS231N). I have a BS in EECS from UC Berkeley, where I worked in Kris Pister's lab.

Email  /  GitHub  /  Google Scholar  /  LinkedIn

Research

I'm interested in computer vision, machine learning, optimization, graphics and robotics.

Flexible Techniques for Differentiable Rendering with 3D Gaussians

We show how shape reconstruction with 3D Gaussians can be expanded to include differentiable optical flow, colored mesh exports and more.

Optimizing Algorithms From Pairwise User Preferences

We show how to perform efficient black-box optimization of algorithm configuration from user preferences. Results include Intel RealSense stereo cameras and a robot social navigation policy.

Discovering Multiple Algorithm Configurations

We show the benefits of discovering an ensemble of configurations for a given algorithm during the course of optimization. Results on stereo, planning and visual odometry.

Fuzzy Metaballs: Approximate Differentiable Rendering with Algebraic Surfaces

An approximate differentiable renderer for a compact, interpretable representation, which we call Fuzzy Metaballs. Our approximate renderer focuses on rendering shapes via depth maps and silhouettes. It sacrifices fidelity for utility, producing fast runtimes and high-quality gradient information that can be used to solve vision tasks.

Venue Analytics: A Simple Alternative to Citation-Based Metrics

A bibliometric/scientometric project. Our main two results are a method of assigning value to all venues in computer science and a method to organize them into distinct subfields, all without using citation data. The resulting venue scores can be used to rank universities’ by scholarly output, and produce a responsive author-level metric.

Direct Fitting of Gaussian Mixture Models

A formulation for fitting Gaussian Mixture Models (GMMs) directly to geometric objects, such as a triangular mesh. This method produces more robust results and produces an improvement in 3D registration for both meshes and RGB-D frames.

Intel RealSense Stereoscopic Depth Cameras

Technical and design details of the Intel RealSense R200 and D400 series

Rigid-body Dynamics for Articulated Mesh Tracking

An invited talk for the HANDS 2015 workshop at CVPR 2015. This includes further details about the efficiency of our rigid-body solver, our machine-learning tools, and some details about our data annotation process.

Dynamics Based Hand Tracking

Using a physics engine (e.g. a dynamics solver) to track 3D articulated objects in real-time.

Dynamics Based Hand Tracking

A tracking algorithm that was real-time on a consumer laptop. Won Best Poster Award.

Intel Projects

Besides my work on the RealSense depth sensors and the publications above, a sampling of my publicly disclosed work

Intel RealSense 400

My responsibilities included system performance, components of the stereo algorithm on the imaging ASIC, and contributions to the design of laser projector pattern.

Compact VCSEL Projector

A low-cost dense, configurable projector system for RGB-D depth sensors.

Depth Image Enhancement

Algorithms to filter, enhance and clean-up RGB-D data streams.

Real-time Box Measurement

Using a single depth sensor, real-time detection of cuboids, accurate estimation of their dimensions, and even some bin-packing.

DashPoint: A low-cost, low-power human interface device

Finger tracking on a microcontroller, with optics tricks and some HCI ideas

Stereoscopic depth reconstruction with probabilistic pixel correspondence search

A fast method for performing stereo depth maps.

Other Projects

These include coursework, side projects and unpublished research work.

Dice Stacking: A Dynamic Manipulation Task

With Hunter Goforth, we designed a manipulation task and solved it with imitation learning.

Introspective Neural Networks

Using pre-trained neural networks to improve fine grained recognition via style transfer.

Stochastic Sampling of Parametric Policies

Using a very simple algorithm to solve some very simple environments

Optimizing for Physical Simulation

With Chris Atkeson and Alex Spitzer. Using optimizers to match an observed trajectory.

A Maze Bot

Making a 6-DoF PUMA arm solve a maze with real-time vision and tracking.

Learning Implicit Communication Strategies

Work with Aaron Goodman on used reinforcement learning to discover implicit collusion strategies in the context of an iterated prisoner’s dilemma.

Computational models for text summarization

Work with Ludwig Schubert on simplified encoders stages for text summarization.

Superresolution Micrscopy

An implementation of Faster STORM using compressed sensing.

Automatically building Restaurant Ontologies

Using the Yelp dataset of reviews to model the semantics and relationships between cuisines, businesses and other properties useful for restaurant recommendations.

Beyond Correlation Networks for the Financial Market

Using graph models, we track the development of financial networks over the 20th century.

Gradient-learned Models for Stereo Matching

Some re-implementations of standard stereo correspondence algorithms, along with experiments using classification for stereo matching.

Multimodal Natural Language Inference

We explored how natural language inference tasks can be augmented with visual data.

CNNs for 3D Model Classification

3D shape classification by learning an embedding function into a 2D image and using a pre-trained ImageNet network. At the time, got state-of-the-art results for single-view classification on ShapeNet40.

Wide-angle Stereo Lenses

We introduce various projection functions in the analysis of stereoscopic depth sensors.

Doctor Bayes

Detecting disease from a short description of symptoms. In some small testing, obtained nearly 90% top 5 accuracy and about 60% top 1 accuracy

Level-set based tracking and segmentation

We implemented a detection and deformable tracking pipeline for red blood cells.

Dequantization of Depth Data

An O(1) time algorithm for producing smooth normals for quantized data, such as the Kinect.

Golf swing monitoring

Work with Ankur Mehta, built a demonstration platform that used wireless low-weight, low-cost sensor platforms to monitor a golf swing.

Project Tetra: Collaborative robot state estimation

With Humphrey Hu, Ryan Julian, and Eric Yuan, a project to show the efficacy of multiple-robot collaborative state estimation. Using Wiimote cameras, mobile robot platforms, and real-time wireless communication.

GINA: Low power design

For testing and validating the functionality of the GINA (Guidance and Inertial NAvigation) mote, a 1.6 gram sensor platform.

GINA: Wireless sensor platform

I helped Anita Flynn and Thomas Watteyne build these small sensors and wrote firmware.

Design and source code from Jon Barron's website