Scribd
Upload
42 views

Confidential 3501132: Appen - Person/Object Segmentation

1. The document provides instructions for segmenting images into pixel-level masks for people and objects. Each person and held object should have its own mask. 2. Rules are provided for mask quality, occlusion, reflections, and definitions for the person and object classes. Examples of proper and rejected images are also shown. 3. Images should be rejected if they have no people, people are too small, extremely blurry, faces are not visible, are black and white, have more than 6 people, have overlays, or are animations or collages. Partial occlusion and slightly blurry images can still be annotated.

Uploaded by

Stephen Velasco Villaruz
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
42 views

Confidential 3501132: Appen - Person/Object Segmentation

1. The document provides instructions for segmenting images into pixel-level masks for people and objects. Each person and held object should have its own mask. 2. Rules are provided for mask quality, occlusion, reflections, and definitions for the person and object classes. Examples of proper and rejected images are also shown. 3. Images should be rejected if they have no people, people are too small, extremely blurry, faces are not visible, are black and white, have more than 6 people, have overlays, or are animations or collages. Partial occlusion and slightly blurry images can still be annotated.

Uploaded by

Stephen Velasco Villaruz
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 16

Appen - Person/Object Segmentation

  Escalation LOG Appen PERSON OBJECT/Feedback

1. High-Level Project Objective

2 l
Objective: To draw pixel-level masks for person & object and mark classifications for each subject based on how they are

13 tia
positioned in the image.  Each class should be instanced; that means, each instance of each of these classes should have their
own mask, even if boundaries of two instances are touching.

IMPORTANT NOTE: 

Subject(s) faces will be blurred out of the image, but please proceed with annotations if any other rejection clauses are
not shown. 

01 n
Consistency: 

e Every person present in a frame must contain a mask and a separate mask if they are holding an object. If
visible, please annotate accordingly. If not visible, please leave the mask empty (add it in the UI matching
the instance of the person you're annotating but don't draw any mask).
To maintain consistency, especially with images with multiple people and objects, please classify the PERSON
annotations from LEFT to RIGHT. 
For the OBJECT class, where 1 PERSON could be associated with multiple OBJECT instances, please annotate from
LEFT to RIGHT (like instructed for the PERSON class), irrespective of who is holding the object.
35 id
Summary:

We segment 2 masking categories:

1. Person (EACH PERSON IN THE OBJECT SHOULD HAVE THEIR OWN MASK)
nf

2. Objects being held by Person(s)

2. Annotation Instructions
RULES

Quality
Co

Please make sure that masks are in the confinement of the individual 
Occlusion
In general, only annotate portions of classes that are visible; we are not to estimate boundaries when portions of
classes are occluded.
Hats can be included in the mask of the person as long as they are not oversized or significant in size. 
Glasses should be annotated over 
Annotate over necklaces and other jewelry
Reflections
Please ignore reflections

CLASS DEFINITIONS

Person
Overall mask of a subject in a frame
For frames with multiple people, please classify as Person1, Person2, etc. so each person(s) has their own unique
mask. 
Object held by Person
Create a detailed mask over any object which is hold/used/played by the person. 
The object size should be relative small or medium respect to the person. 
Holdable: Camera, Cup, Bottle, Toy
Pets: Dog, Cat, Rabbit, Hamster
Instrument: Guitar, Violin. 
Fashion Items: Bag
Not to draw the mask. 

2 l
Relative Large: Desk. 
Examples:

13 tia
01 ne
35 id
nf
Co

Visual Examples
Description: Focus on the visible part. For the “person” segmentation, the segmentation only includes the visible region.
2 l
13 tia
Description: If the person holds more than an object, please label them as the different classes, such as object-00, object-01.
The ordering here does not matter. 

Rejections
WHEN TO REJECT IMAGES:

01 neNo people in an image


People are too small (e.g., area is less than 5% of the image)
People are extremely blurry (e.g., motion blur). And it’s too hard to draw the mask. 
Subject(s)' face isn't / aren't visible
Black-and-white or grayscale images
35 id
There are more than 6 in the image
Artificial overlays (e.g. stickers, face filters from FB Messenger or Snapchat)
IMPORTANT NOTE: SUBJECTS FACES WILL BE BLURRED FROM THE IMAGE AND THIS DOES
NOT COUNT AS A REJECTION FOR ARTIFICIAL OVERLAY
Animation images. 
Photo collages
nf

Person is in the other object. (Painting, book cover)


Unnatural borders on the image
“Blurred area significantly prevents me from drawing high quality mask along important boundary”
IMPORTANT: if there are multiple people and you feel that there are subjects that cannot be masked with high
quality die to the blurred area, REJECT. 
Co

REJECTED EXAMPLES:
2 l
13 tia
01 ne
35 id
nf
Co

“Blurred area significantly prevents me from drawing high quality mask along important boundary”
Co
nf
35 id
01 ne
13 tia
2 l
Collages
Co
nf
35 id
01 ne
13 tia
2 l
2 l
13 tia
01 n
Extreme Blur (The person in the left side)
e
35 id
nf
Co

Animated Object
Co
nf
35 id
01 ne
13 tia
2 l
Artificial Overlay

2 l
13 tia
01 ne
35 id
nf
Co

Person in a photo
Co
nf
35 id
01 ne
13 tia
2 l
2 l
13 tia
01 ne
35 id
nf

5. Con icted Examples


PEOPLE OCCLUDED BY COMPLEX OBJECTS

If excluding an object will make the mask very complex, you can include the object as part of the body (wire netting is
Co

included in the example).

BLURRY PEOPLE

Please annotate the images where people are slightly blurry (the three examples below should be annotated). Only reject the
images if people are extremely blurry, so much that you cannot annotate them.
2 l
BOUNDARIES OF BLURRY PARTS

For the boundaries, include just the solid parts and exclude blurry parts (the parts that are slightly transparent where the

13 tia
background can be seen). See the hands in the example.

01 ne
HALO TOOLS/TIPS: 
35 id
nf
Co
Co
nf
35 id
01 ne
13 tia
2 l
Co
nf
35 id
01 ne
13 tia
2 l
Co
nf
35 id
01 ne
13 tia
2 l
Co
nf
35 id
01 ne
13 tia
2 l

You might also like