[visionlist] [meetings] RSS 2020 Tutorial on "Certifiable Robot Perception: from Global Optimization to Safer Robots"

[Tat-Jun Chin]

Dear colleagues,

We cordially invite you to attend the RSS 2020 tutorial on “Certifiable
Robot Perception: from Global Optimization to Safer Robots”.
We will discuss novel optimization tools for robust 3D robot perception,
provide a snapshot of the state of the art, and outline a number of open
problems.
We trust this will be a useful event, with multiple talks, hands-on
examples, and open discussion. Please find more details below:
-----------------------------------------------------------------------------------------------------------------------------------------------------
Certifiable Robot Perception: from Global Optimization to Safer Robots
(tutorial)
Robotics: Science and Systems (RSS)
Date: July 12, 2020
Location: Zoom - Due to COVID-19, this tutorial will be virtual and
accessible via the RSS portal (Session WS1-3).
Website: https://mit-spark.github.io/CertifiablePerception-RSS2020/

OVERVIEW:
----------------
This tutorial gives an in-depth introduction to global optimization tools,
including convex and semidefinite relaxations, applied to robot perception
problems. The first goal of the tutorial is to motivate the need for global
solvers by providing real-world examples where the lack of robustness
results from the difficulty in solving large optimization problems to
optimality. The second goal is to provide the attendees with basic
mathematical and algorithmic concepts, and survey important recent advances
in the area. The third goal is to outline several open research avenues:
global optimization has an enormous untapped potential and it is hoped that
this tutorial will inspire researchers to use modern optimization tools to
solve several outstanding challenges in geometric robot perception. This
tutorial aims to replicate the success of the “twin” tutorial “Global
Optimization for Geometric Understanding with Provable Guarantees” held at
ICCV’19 .

DETAILED DESCRIPTION:
------------------------------------
Understanding the geometry of a scene from sensor data is a key requirement
from many applications, ranging from autonomous vehicles (e.g.,
self-driving cars, consumer drones) to Virtual and Augmented Reality.
Indeed, geometric understanding encompasses several core topics in computer
vision and robotics, including Structure from Motion and SLAM, point cloud
registration and object pose estimation, single-view and multiple-view
geometry, among many others. Despite the maturity of the algorithms
developed for geometric understanding, both researchers and practitioners
are well aware that the presence of large noise, outliers, and missing data
makes modern pipelines brittle when operating in the wild. An incorrect
understanding of the geometry of the scene can negatively affect the user
experience in consumer applications, and may put human life at risk in
safety-critical applications, such as self-driving cars.

One of the main causes of the fragility of geometric understanding is the
fact that modern pipelines typically trade-off performance (robustness,
optimality, guarantees) for computational efficiency. Indeed, many
optimization problems underlying geometric understanding are intractable
due to non-convexity or due to their combinatorial nature (e.g.,
measurement selection to reject potential outliers). Several algorithms
obtain fast solutions using local iterative nonlinear solvers or heuristic
approaches. While heuristics and local solvers tend to work well in the
low-noise and low-outlier regime, they are prone to fail in more
challenging real-world conditions.

Global optimization and convex relaxation techniques have been recently
shown to provide an effective tool to tame the complexity of geometric
understanding while enabling efficient solutions. A sequence of papers have
debunked two common misconceptions behind these methods. The first
misconception is that a convex relaxation always entails computing an
“approximate solution”: indeed recent papers have shown that, in certain
regimes, one can design convex relaxations that (provably) solve the
original problem exactly. The second misconception is that many of these
relaxations, in particular semidefinite relaxations, are slow in practice:
modern solvers and theory instead show that one can solve several classes
of such relaxations very efficiently. A growing body of work in vision and
robotics demonstrates that these tools can indeed solve in a provably
optimal manner problems including SLAM, 3D registration and object pose
estimation, as well as combinatorial problems involving outlier rejection.

This tutorial aims to give an in-depth introduction to global optimization
tools, including convex and semidefinite relaxations. The first goal of the
tutorial is to motivate the need for global solvers by providing real-world
examples where the lack of robustness results from the difficulty in
solving large optimization problems to optimality. The second goal is to
provide the attendees with basic mathematical and algorithmic concepts, and
survey important recent advances in the area. The third goal is to outline
several open research avenues: global optimization has an enormous untapped
potential and it is hoped that this tutorial will inspire researchers to
use modern optimization tools to solve several outstanding challenges in
geometric vision.

ORGANIZING COMMITTEE:
-------------------------------------
- Luca Carlone, Massachusetts Institute of Technology
- Tat-Jun Chin, The University of Adelaide
- Anders Eriksson, University of Queensland
- Heng Yang, Massachusetts Institute of Technology

FURTHER INFORMATION:
-----------------------------------
Please send any questions to Luca Carlone (lcarlone at mit.edu), Tat-Jun Chin (
tat-jun.chin at adelaide.edu.au), Heng Yang (hankyang at mit.edu), or Anders
Eriksson (a.eriksson at uq.edu.au).

All the best,
Luca (on behalf of the organizers)
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://visionscience.com/pipermail/visionlist_visionscience.com/attachments/20200703/fa636d9b/attachment.html>