Google Docs를 사용하여 게시됨 악용사례 신고 자세히 알아보기

(EXTERNAL) Instructions - Anteater 3D TSR 3FPS, FWP - v5.0 5분마다 자동으로 업데이트됨

Anteater 3D Traf몭c
Signs Recognition: Spec
doc
LIVE

Instructions
­ Anteater 3D
TSR 3FPS,
FWP
Last update: Sep 25, 2024

­PLEASE READ
CAREFULLY­
READ THE FOLLOWING
UPDATES:

Relevance is now an
attribute of cuboids, not
projections. For this reason,
it should be annotated
throughout the entire clip (if
applies).

New updates: If the cuboid

is in front or behind of
the ego vehicle and is
more than 100
meters away, its relevance
is 'other.'

if the sign does not

generate projections its
relevance is ' other'

Reminder:
If no projections are
available in any cuboid,
then you must delete it,
otherwise you’ll be
blocked with an error in
the Quality Assistant.
 ­PLEASE READ
CAREFULLY­

Key Sections

● Summary of task
● Workflow
● When to annotate objects?
● Annotation Rules

● Table of Labels & Attributes

Taxonomy
● Label De몭nitions
● Attribute De몭nitions
● Annotation Tips

Summary of task:

❖ In anteater TSR 3D Phase 2

you will annotate all visible
Traffic Signs and assign them
attributes.

❖ TSR tasks have around 300

frames. You should look at
every frame carefully to find
all traffic signs in the task.

❖ Please create cuboids for

each object in the task at one
single frame, each object
should be tracked with a
single ID.

❖ Please note that in this new

version of tasks, we will
encounter some shifting in
annotations. We have
designed this new workflow to
deal with shifting data. What
is important is that a cuboid is
well fit to an object at a single
frame, ideally near the object
and with a clear view of it in
the camera. This single frame
is the “reference frame.”

❖ All objects in this task are

stationary and should apply
for all frames.
❖ Main Sign: A single traffic sign
with clear standalone
instructions. Can be
embedded in a signboard.
❖ Ego Car: The car that
collected the data, i.e. the car
on which the cameras are
mounted.

❖ All Traffic Signs in the frames

of interest must be labeled
without exception. This
includes the back of traffic
signs.

Workflow

Please follow steps to successfully

label TSR tasks with high quality:

1. Scroll through the frames to

familiarize yourself with the
task, take note of areas that
are congested with objects or
have objects that are far
away/hard to see.
2. Annotate the scene
annotation "
Is_good_data " If the
attribute is "Yes" processed
to look for a sign to start
annotating.
3. Identifying a Reference
Frame
a. To start, do a quick
scan for the best
frame for a traffic sign
and draw a rough
cuboid around it.
i. Use dynamic
APC to get the
rough cuboid.
ii. Enable Partial
APC size 3

b. Scroll through the

frames in the camera
view to find the frame
that clearly displays
the traffic sign.
i. The ideal
frame to check
should be the
frame where
the sign is
greater than or
equal to 20
meters away
(i.e., the first
frame where
the distance is
at least 20
meters from
ego) AND the
object is not
occluded so
we get good
LiDAR density.
1. You
can
confirm
good
LIDAR
density
if the
points
cover
the full
height
of the
sign in
the
chosen
frame
ii. If there is any
ambiguity
between which
two frames to
choose, select
the frame
where the sign
is closer to 20
meters from
ego.
c. Note: You may need
to change the
keyframe for a static
cuboid. To do this,
convert the cuboid to
dynamic, add a
keyframe at the new
reference frame, and
then convert the
cuboid back to static.
This will CLEAR the
attributes if they are
already set

4. Starting with the first sign,

working in an order where the
nearest signs are labeled first,
select a sign to label.
a) Size the
cuboid
 cuboid
according to
the LiDAR
points for the
reference
frame using
partial APC
with a size of 3
frames.
i)
IMPORTANT:
Note
that the
cuboids
will not
be
perfectly
tracking
the
LiDAR
points
for the
other
frames
due to
shifting/drifting
data.
But it is
completely
fine –
this is
expected.
What is
important
is that
the
cuboid
is well
fit at
the
frame
of
reference.

b) Draw the first

label on the
first frame,
make sure the
label is tight to
the visible
sections and
assume the
entire size of
the sign, if any
portion of it is
occluded.

i) Use pitch &

roll to ensure
the cuboid
covers the
LiDAR points
perfectly on
the reference
frame with
partial APC
enabled.

c) In the event of
blooming at
the reference
frame, use the
density mode
for the points.
i) If
density
mode
is not
useful
(i.e.,
severe
blooming),
then
find the
frame
that is
nearest
to the
object
where
the
object
can be
clearly
made
out.
5. After annotating the object
a. For the following
attributes (IRF,
Readability), we
MUST turn on
Projection Mode
(click ‘J’ while in
camera mode) to be
able to annotate said
attributes. We will only
be able to annotate in
frames with
projections.
 b. You can confirm
projection mode is on
by checking the
bottom left corner of
the camera view, the
words “projection
mode” should appear,
highlighted in yellow.
c. The cuboids will also
appear like this:

d. Once projection is
mode is on, you will
be able to annotate
readability and IRF
e. P.S DON’T double
click any attribute to
set it across all
frames, please set the
attributes in EACH
INDIVIDUAL
AVAILABLE FRAME

6. Move to the next frame that

an unlabeled object appears
and repeat all steps.

I. Edge Cases (PLEASE

READ)
A. When an object
cannot be viewed in
the camera with
sufficient lidar density
(e.g., tall objects at the
beginning of a scene),
then select the frame
nearest to the ideal
distance where the
object is clearly visible
in lidar after you pass
the object.
B. When an object never
has an unoccluded
view using 3 frames
APC (e.g., following a
large vehicle), then
incrementally increase
the number of frames
for APC by steps of 3
to find a frame with
sufficient lidar density.
C. When an object does
not have an
unoccluded view near
the object but has a
good view at
significantly further
distances (e.g.,
100m), then use the
further frame with
good lidar density.

When to annotate
objects?

One things need to be

considered when trying to
annotate in this project:
 1. Is it visible from the

camera?

If YES, then ANNOTATE THE OBJECT!

If NO, we DO NOT ANNOTATE THE

OBJECT!

Note: Once a sign has been annotated

based on the criteria, it shall continue to

be annotated in the following frames.

Example 1:
Is it visible from the camera? Yes
CONCLUSION: ANNOTATE

Example 2:
Is it visible from the camera? No
CONCLUSION: DO NOT ANNOTATE

Example 3:
Is it visible from the camera? Yes
Does it have the minimum lidar points? no
CONCLUSION: ANNOTATE
Annotation rules:

Rule Example

Each camera
recording shall be
annotated from start
frame to end frame
with a frequency of 3
camera frame per
second
(corresponding to
every 3rd LiDAR
frame)

Each annotated
frame shall have all
visible traf몭c signs
fully annotated
according to the
annotation
instructions.

In case no signs are

visible in a speci몭c
frame that has been
considered for
annotation, that
frame shall still be
marked as annotated
for
TSR.

In a sequence of
consecutive
annotated recordings
a physical sign shall
be assigned the
same ObjectId
across recordings.

Annotation of
supplementary
signs
All annotated main
signs and
signboards shall
reference its relevant
supplementary
signs.
If a supplementary
sign is relevant for a
signboard and all its
embedded main
signs, it shall be
referenced by the
signboard alone.

The embedded signs

shall be annotated
though, but not
referenced from the
signboard

Supplementary signs
shall be annotated in
the same manner as
main signs.

Annotation of
embedded signs
Embedded main
signs shall be
annotated in the
same manner as
stand alone main
signs.

Annotation of
Relevance
Whether the sign is
relevant or not for
any road user at the
position of the ego
vehicle

In the current
frame. Conditional
restrictions
on signs, such as
vehicle type
restriction, time
restriction, distance
restrictions, weather
restrictions etc, shall
not be
considered.
For example a sign
that is relevant for
the ego vehicle at a
speci몭c time should
be considered
relevant even if the
current time does
not match the
restriction

Heading for cuboids

The orientation
should be where the
signal is pointing.
Global Question-=

Is_good_data

Table of Labels & Attributes

Taxonomy:

Label Attributes

Main Traf몭c Sign Readability

Relevance

Signboard Readability

Relevance

Supplementary Readability
sign
Relevance

Linkedtosign

_________________________________________________________________________________________________________________
Label Definitions

Label De몭nition Example

A single traf몭c
sign with clear
standalone
Main Traf몭c instructions.
Sign
Clari몭cation:

Each sign
should have its
individual
bounding box.

There can be
more than one
main sign on
the same pole.

ATTENTION SIGNS must be annotated as

Main Traf몭c Signs

IMPORTANT NOTE:
United States Exit Signs must be labeled as
Main Traf몭c Signs, this exception only applies for
this country, each of them should be labeled
individually as Main Traf몭c Sign.
 For Example:

A traf몭c sign
that contains
Signboard one or more
embedded
signs or
supplementary
signs.

The signboard
is the entire big
sign.

You MUST see

signs inside to
use the
Signboard
label ( even
though its big,
or square, if
you don’t see
signs inside DO
NOT label it as
a Signboard

Never use
Signboard for
back signs
(you can’t
con몭rm there’s
traf몭c signs
inside it)

A traf몭c sign
Supplementary that contains
information
that alters the
interpretation
of a nearby
main sign or
signboard.
Note: the
supplementary
sign must be
separate from
the main sign
(Case A)or
delimited by a
border (Case
B)

Embedded = Traf몭c sign

MTS contained
inside another
traf몭c sign
(signboard).

It’s a traf몭c
sign
embedded in a
signboard.

Please look for

these extra
carefully
Only label the
traf몭c sign
inside as
embedded.

The full traf몭c

sign in which it
is mounted
should have a
 should have a
separate
annotation as a
signboard.
Usually they
don’t appear
alone or
isolated.

Attributes Definitions

Relevance

Whether the sign is relevant or not

for any road user at the position of
the ego vehicle
● Look carefully at the image
and select the correct option
for this attribute.

Option De몭nition Example

Relevant Sign is relevant.

Not considering
sign
restrictions and
that ego vehicle
will continue to
drive in the most
probable path,
often ego lane.

Signs marked in green are relevant

RelevantNeighboringLanes Sign is not relevant

for the ego vehicle
lane, and directly
affects one or
more neighboring
lanes in the same
driving
direction. The
neighboring lane
must be on the
same road and not
separated by a
solid double lane
or physical barrier.
 solid double lane
or physical barrier.
It must be parallel

Signs marked in yellow are RelevantNeighboringLans

RelevantExitLane Sign is not relevant

for the ego vehicle
but is relevant for
the exit lane.
Only applies when
there is no
physical barrier or
there is a
possibility for the
ego vehicle to take
the exit, but we
know it won't.
Note that if the ego
vehicle is driving in
the exit lane the
sign should be
annotated as
relevant.
It is only valid for
parallel lanes that
exit a motorway.
If the sign is
beyond the point
where the excite
lane diverges from
the motorway, it
should be
annotated as
"other" until the
frame where the
ego vehicle has
passed the point
where the lane has
diverged. After
that point, it
should be
annotated as
"NotRelevantOther
Road"

NotRelevantOtherRoad Sign is not relevant Signs marked in red are not relevant
for the ego vehicle
but is relevant for
a road that is not
reachable from
the current ego
vehicle
position for
example a parallel
road. A road is
considered non-
reachable if the
ego vehicle needs
to cross a solid
double lane or a
road boundary in
order to reach it.
There must be a
traf몭c impediment
or some physical
barrier that
prevents the ego
vehicle from
reaching that road.

Unknown Sign relevance

can’t be
determined.
100% occluded
signs and within
the 100 meter
range

Used for all objects

located more
Other - meters away
from the vehicle
(X>100m),
regardless of
whether the signal
is occluded or
not,in front or
behind.
Behind the ego vehicle
To determine this
distance use the
“Distance to the
device” value in
the menu.
 When the cuboid
does not generate
projections.
When the ego
vehicle has already
passed the signal
and it is now
behind the ego
vehicle. This case
only applies to
signs that are Signs on busses are other
initially in front of
the vehicle. If the
sign is behind the
vehicle from the
start of the clip, it
should not be
annotated.

Used for all signs

that are located on
connecting
roads to
intersections or
roundabouts --
All Not Readable
Back of Sign shall
be annotated as
other throughout
the entire clip.
Truck signs and
traf몭c sign
mounted on back
of trucks, walls or
other objects
should be
annotated as
"other", unless it is
clear that they are
relevant for the
ego vehicle.
Used for all other
cases where any of
the above values
cannot be used.

Readability

Whether it is possible to distinguish

the traf몭c sign type or not in the
current frame.
This attribute determines how
readable the information of a signal
is in the current frame.
UPDATE! THIS SHOULD ONLY
BE SET EVERY 3 LIDAR
FRAMES. (1,4,7,10, etc.)
This applies for attempters and
reviewers.
Please Only review and set the
attribute every 3 lidar frames + the
last one. (The frames in the middle
should be ignored/not looked at/not
reviewed)
Only annotate and review this
attribute in the following frames
1,4,7,10,13,16,19,22…
The other frames do not review
them, leave the automated value

Option De몭nition Example

Con몭dent It is
possible to
see the
sign type
without
any
dif몭culties

Not It is not
Readable possible to
know
exactly
which sign
type is in
the current
frame

Not It is not
Readable possible to
LED uniquely
Fragmented distinguish
the sign
type in the
current
frame due
to LED
issues

Not It is not
Readable possible to
LED uniquely
 Readable possible to
LED uniquely
Inactive distinguish
the sign
type in the
current
frame as
the LEDs
are
inactive

Not The traf몭c

Readable sign is
Back of facing
Sign away from
the
camera in
the current
frame

Procedure: Look carefully at

the image and select the
correct option for this
attribute.
Note: The zoom function can be
used to determine readability.
For a signal to be "Con몭dent" the
information in the signal must be
interpretable.
If the information is not visible or
the image quality is so blurred as
to prevent interpretation of the
signal, the correct answer is "Not
Readable"
If only the back of the sign is
visible then the readability is
"Not Readable Back of Sign"
Take the following examples into
consideration:

VISUAL EXAMPLES ON
READABILITY ATTRIBUTE:
Readability Examples

Con몭dent Not Readable

Con몭dent Con몭dent Con몭dent
Con몭dent Not Readable

Con몭dent Con몭dent
Con몭dent Con몭dent
Con몭dent Not Readable Not Readable

Not Readable Not Readable

Not readable Not readable
Con몭dent Con몭dent Not readable

VISUAL EXAMPLES ON
READABILITY ATTRIBUTE:
Readability Examples

Con몭dent Con몭dent
Not Readable Not Readable

Con몭dent Not Readable Not Readable

Con몭dent
Con몭dent Not Readable Not Readable
Con몭dent
VISUAL EXAMPLES ON
READABILITY ATTRIBUTE:
Readability Examples

Con몭dent Not Readable Not Readable Not Readable

Con몭dent Con몭dent
Con몭dent
Con몭dent

Con몭dent Con몭dent Not Readable Not Readable

Con몭dent Con몭dent Not readable
Not readable

Con몭dent Not Readable Not Readable

Con몭dent Con몭dent Con몭dent Not Readable Not Readable

_________________________________________________________________________________________________________________

Attribute Example

LinkedtoSign Main sign with

To which main supplementary
sign is this sign Sign
linked to.

This attribute only

applies for
Supplementary
Sign

Sign 2 is a Supplementary Sign

from Main Sign 1, and it should
be linked to it.

Attribute Option De몭nition Example

yes should be If the object
good when looks correct
Is_good_data data is good. (annotatable) we
It might annotate the
have a bit of task - select yes
shifting data
but it's still
annotatable,
drift is
small.
 No There is a If the object is
lot of drift, very
data is bad, distorted/deform
and hard to that is very hard
annotate to annotate
select - no and
do not annotate
the task

Annotation Tips

Next you will find some tips and

examples on how to properly
annotate the different types of signs
in these tasks:

Label De몭nition Example

Single main
sign

Main sign with 1. Bounding box

supplementary around the
sign main sign

2. Bounding box
around the
supplementary
sign

● Supplementary
signs
referenced to
from the main
sign

Mutually 1. Bounding boxes around

exclusive main main signs
signs with one
2. Bounding box around the
supplementary
supplementary sign -
sign
● Supplementary signs
referenced to from the
closest main signs above

Main signs 1. Bounding

with shared boxes around
supplementary main signs
sign
2. Bounding box
around the
supplementary
sign

● Supplementary signs
referenced to from
both main signs

Main signs 1. Both

with shared signboards are
supplementary annotated
sign
supplementary annotated
sign
2. Bounding
boxes around
main signs

3. Bounding box
around the
supplementary
sign

4. Bounding box
around the
supplementary
sign
==
● Supplementary
Signs referenced to
from the lower
signboard.

Invalidated
main and 1. Bounding box
supplementary around the
sign main sign

2. Bounding box
around the
supplementary
sign

● Supplementary
Signs referenced to from the
main sign

Invalidated Shall not be

sign, annotated
completely
covered

Zone End sign 1. Bounding box around

with the signboard
embedded
signs 2. Bounding box around
all the embedded
main signs

3. Bounding box around

the embedded
supplementary signs

● Supplementary
signs
referenced to
from main sign
above them

Living street 1. Bounding box around

sign with the signboard
embedded 2. Bounding box around
speed limit all the embedded
sign main signs

Note: Set the size of

the bounding box
around the
embedded main sign
to the expected full
size of the main sign.
City limit sign 1. Bounding box around
with the yellow signboard
embedded
speed limit 2. Bounding box around
sign the embedded main
sign -

3. Bounding box around

the white main sign

Main sign 1. Bounding box

including around the
distance whole main
information sign

Conditional 1. Bounding box

signboard around the
signboard

2. Bounding box
around all the
embedded
main signs

3. Bounding box
around the
embedded
supplementary
sign

● Supplementary Sign
referenced to from
lower main sign

Multilane 1. Bounding box

signboard around the
Signboard

2. Bounding box
around all the
embedded
main signs

3. Bounding box
around the
supplementary
sign

● Supplementary
sign
referenced to
from signboard

Signboard 1. Bounding box

with around the
embedded Signboard
main signs
2. Bounding box
around all the
embedded
main signs
US exit main 1. Bounding
signs boxes around
each main sign

Large main 1. Bounding

signs boxes around
large main

sign

2. Bounding box
around small
main sign on
top

Signboard
with
supplementary
sign
1. Bounding box
around the
Signboard

2. Bounding box
around the
embedded
main sign

3. Bounding box
around
embedded
supplementary
sign

● Supplementary
signs
referenced to
from signboard

Traf몭c signs 1. Bounding box

not directly around the
relevant for signboard
ego vehicle
2. Bounding box
around the
embedded
main sign
Back of signs 1. Bounding box
around all
main signs

2. Bounding box
around all
main signs

Sign partly 1. Bounding box

outside of around the
image estimated
position of the
complete sign

Version 1.0 ­ 10/23/2023