SotM 2017
Deriving map data from street-level images
Slide 2
Slide 2 text
15:15 | Mapillary recap - where are we now
15:20 | Computer vision developments
15:40 | Community stories
15:50 | Mapillary tools
16:00 | Map editing from street-level imagery
16:30 | Fin
Agenda
Slide 3
Slide 3 text
Understand how
street-level imagery
can be applied in
mapping
Inspired by other
community projects
Aware of available
tools
Objectives
Widen scope of map
editing to include new/
ignored attributes
Engage others in your
locality to contribute
street-level imagery
New-timers Old-timers
Slide 4
Slide 4 text
172,373,552 images, 3,141,352,533 meters
Slide 5
Slide 5 text
Hardware
Slide 6
Slide 6 text
State of the apps (iOS & Android)
Action cam setup
360º cams - are they any good?
Professional (Ladybug)
Hardware
Slide 7
Slide 7 text
Mobile Apps
Slide 8
Slide 8 text
360º cameras
Slide 9
Slide 9 text
Semi-professional
Slide 10
Slide 10 text
The pro’s
Slide 11
Slide 11 text
Images by device
42.7%
25.0%
15.7%
12.1%
3.9% 0.7%
Android Action cam iOS Other Professional Windows
Slide 12
Slide 12 text
Computer vision
Slide 13
Slide 13 text
Privacy
Slide 14
Slide 14 text
Blurring
License plate blurring
Face blurring
Slide 15
Slide 15 text
Point clouds
Slide 16
Slide 16 text
The method we use to create a
3D environment from images
Points are identified across
images and combined with
camera data such as GPS
information and angle
Image locations can be
improved with this new
information
3D models improve viewing
experience and object detection/
positioning
Structure from Motion
Slide 17
Slide 17 text
Semantic segmentation
Slide 18
Slide 18 text
25,000 human annotated images
Geographically diverse
Diversity of cameras and conditions
100 object categories, 60 instance specific
Training set for our deep learning models
Free for research purposes
Mapillary Vistas Dataset
Slide 19
Slide 19 text
Semantic Segmentation
Slide 20
Slide 20 text
No content
Slide 21
Slide 21 text
AI Detections
Slide 22
Slide 22 text
Point cloud to map data
Point cloud data with semantic classifications
Slide 23
Slide 23 text
Point cloud to map data
Map data overlaid
Slide 24
Slide 24 text
Point cloud to map data
Map data extracted
Slide 25
Slide 25 text
Community stories
Slide 26
Slide 26 text
Bangladesh - MapDhaka
Slide 27
Slide 27 text
Bike Ottawa
Slide 28
Slide 28 text
Bike Ottawa
Slide 29
Slide 29 text
Fukushima
Slide 30
Slide 30 text
Urban mobility
Slide 31
Slide 31 text
University of
Washington
Capstone Project
Slide 32
Slide 32 text
University of Washington
Slide 33
Slide 33 text
University of Washington
Slide 34
Slide 34 text
University of Washington
Slide 35
Slide 35 text
Mapillary tools
Slide 36
Slide 36 text
V2 has been deprecated
GeoJSON default geographic format
No more search in API - v2/search/im > v3/images
Leaderboards built in
Easier bounding boxes
API V3
Slide 37
Slide 37 text
OSMAnd
Slide 38
Slide 38 text
#CompletetheMap
Slide 39
Slide 39 text
Map editing
Slide 40
Slide 40 text
1. Get out our laptop and a mouse and open
the editor of your choice. iD editor is
recommended because it supports 360 degree
imagery, something JOSM is not currently
capable of. If you prefer JOSM, you may wish
to open a Mapillary windows separately.
Step 1
Slide 41
Slide 41 text
Step 1
Slide 42
Slide 42 text
2. Choose an area you would like to map.
• The area around Aizuwakamatsu City Cultural
Centre should hopefully have a lot of imagery from
our photo walk and the awesome Fukushima
mapping community.
• Amsterdam is a fun place to start because there are
800,000 high quality panoramic images to work with.
Step 2a
Slide 43
Slide 43 text
2. Choose an area you would like to map.
• MapLesotho continues to improve. Help them add greater detail in Maseru
and elsewhere using imagery captured from a range of devices.
• Choose an area you’re more familiar with and take a look at what Mapillary
images are available in the area.
• You can do this in iD editor by clicking Map Data and enabling the
Photo Overlay.
• Get familiar with objects that Mapillary automatically detects in images
and see if you can use them to edit. Access them via mapillary.com by
selecting AI Detections up the top.
Step 2b
Slide 44
Slide 44 text
3. Work through the list of tags you can add.
• Try editing some features you wouldn’t usually.
• Use street-level imagery or detections as your primary
data sources.
• If you're uncertain about any terms, the OSM wiki is a
great resource, as are the mappers sitting next to you.
Step 3
Slide 45
Slide 45 text
4. The internet cuts out at 16:30 sharp. Make sure
you have added changeset comments and saved
any edits you’re currently working on before then.
Step 4
Slide 46
Slide 46 text
Buildings
Slide 47
Slide 47 text
Roads
Slide 48
Slide 48 text
Ammenities & nature
Slide 49
Slide 49 text
Parks
Slide 50
Slide 50 text
Remember to add your changes comments and
save
Thanks for editing
Slide 51
Slide 51 text
170 million + images. Let’s make the most of them
Different devices being used to contribute
Technology allowing us to derive more map data
Community stories & local inspiration
Additional ways you can edit with street-level
imagery
What have we covered?
Slide 52
Slide 52 text
Map with us!
ed@mapillary.com
hello@mapillary.com
@mapillary
@mapillary