|
1. Label as
many objects and regions as you can within the same image.
Trace the boundary of an
object.
It is better to label several
objects from the same image, than to label one object in many images.
Then you will be asked
to enter its name (e.g., car, building, etc.). Use a name that you think
other people would also use.
You can use multiple words to describe one object.
|
2. Follow the object boundary, ignoring
occlusions.
When there are
occlusions, follow the boundary of the object as if it was not occluded:
 
If you label multiple
objects, we can later reason about which parts of the image are
occluded. We know the car should be
on top of the road:
 
If an object is heavily
occluded, then you only need to label the visible region.
|
3. Label regions,
objects and parts
We are interested in
objects such as cars, pedestrians, tables. But we are also interested in
regions such as the sky, buildings, sidewalks, walls, etc.
You can also label parts
(e.g., the legs of a table, the wheels of a car).
We can use that
information later to reason about what objects are part of others by
studying how many times they overlap.
|
