The image shows a scientific poster presentation titled "An N-Point Linear Solver for Line and Motion Estimation," authored by Ling Gao, Daniel Gehrig, Hang Su, and Davide Scaramuzza. The research is supported by the ShanghaiTech University and the University of Zurich. 

The poster is divided into several sections:

1. **Motivation**: Describes the need for recovering partial linear velocity and line parameters from minimal data, driven by the advancements in 3D line and gyro readings using IMU.

2. **Contributions**: Lists the main contributions of the study, including a novel solver for minimal and overdetermined systems, an a priori usage of batch optimization, and a new velocity averaging scheme for accurate velocity results.

3. **Methodology**: Outlines the process that begins with extracting events from a temporal slice, followed by linear solver application, and culminating in average partial velocity results.

4. **Explanation of Terms**: Provides a succinct definition and explanation of "What is an event camera?" which is pivotal for the context of the study.

5. **Multiple Solutions**: Discusses how the proposed solver can return up to 4 different solutions, which correspond to symmetrical transformations.

6. **Characterizing Event Manifolds**: Explores the mapping of event manifolds and their representation.

7. **Quantitative Results**: Shows numerical results and graphs demonstrating the accuracy and efficacy of the proposed linear solver and velocity averaging scheme across various complexities.

The poster includes various illustrative graphs, mathematical formulas, and diagrams to convey intricate details of the methodology and results to the audience. Two people are partially visible on the right and left edges, suggesting an interactive presentation setting.
Text transcribed from the image:
上海科技大学
ShanghaiTech University
University of
Zurich
Motivation
Highli
An N-Point Linear Solver for Line and Motion Esti
Use events generated by a 3D line and gyro. readings from IMU to
recover partial linear velocity and line parameters with a fast and
robust linear solver. Applying a novel velocity averaging scheme, we
fuse these partial observations to obtain full linear camera velocity.
Contributions
1. A linear solver for minimal and overdetermined (N>=5 events)
systems, that is 600x faster than polynomial solvers.
2. A 3 DoF angle-axis-based line parametrization that improves the
numerical stability of existing solvers.
3. A full characterization of degeneracies and solutions of the solver,
and manifolds spanned by the events.
4. A geometry-inspired velocity averaging scheme that is simpler and
faster than existing method.
What is an event camera?
Ling Gao*, Daniel Gehrig*, Hang Su, Davide Scaramu
Methodology:
Events from
temporal slice
Linear Solver
Stack linear incidence
relations, one for each event
Solve for line and velocity
unknowns using SVD
Reconstruct the unknowns
using simple vector algebra
Incidence Relationship
Extract manifold
*indicates equal contribution
Average partial
velocity results
linear velocity projection
Velocity Averaging Scheme
3D line params.
L-eee
•Stack linear velocity constraint, one for each line
• Solve for full velocity using SVD
P
rotated event
bearing vector
camera C
frame
e
e
3D line
Partial Velocity Measurements
m
R
Standard
camera
output
0000000
stock
-
camera
linear velocity y
origin
unknowns
simplify
Unobservable
due to
ef ((Rut))+(-e) = 0
=>
f(ue-use) += 0
measurements
aperture problem
line index
RR
=0
last column of V
from SVD/A)
=>
e
=
v = + +
geometric
constraint
=>
v (e
event
camera
output
no motion,
no events!
reconstruct
unknowns
sures a stream of asynchronous brightness changes ("events")
antages: high temporal resolution, reduced motion blur, low
er consumption, high pixel bandwidth, high dynamic range
tiple Solutions
roposed solver returns up to 4 different solutions
symmetry corresponds to flipping along the z-axis
The second corresponds to flipping the line direction
• Disambiguate by checking the line is "in front of the camera"
AER
e=246
u = 46
21:3×24:6
multiple
solutions
collect
variables
->
Characterizing Event Manifolds
flipping
flipping
line direction
flipping
line direction
flipping
z-axis
-Bum
A
de-rotate events with
rotate into
canonical frame
IMU angular rates
All event manifolds can be mapped into
a canonical representation in frame R
In this frame, the manifold is only
parametrized by u, u
u controls the curvature, while u
controls the slope
<o
project onto
y-t plane
10
>0
= [213] - [ue - ue]
multiple solutions
Solution
R
[ee] 0 = (log R₂)
u = [0 2 น.]
stack
last column of V
unen-une
Soluti
from SVD(A)
v=0
v = +
AERNX3
Quantitative Results
Our linear solver and velocity averaging sch
can be extended to an arbitrary number of ev
Pixel Noise (0.5 px)
Time. Jitter (0.5 ms)
Gyro. Noise (5.0°/s)
6 7 8
9 10 20 30 40
Number of used events
50 100 1000
As few as 10 events are sufficient to
substantially reduce the error from
noisy measurements.
Direction Emm
• As
e