RGB-D Object Recognition for Semantic Navigation

RGB-D Object Recognition for
Semantic Navigation
ENSTA ParisTech INRIA FLOWERS Team
Louis-Charles Caron
David Filliat
What is semantic navigation ?
Using high-level “semantic” information to navigate
ENSTA ParisTech
Thème Robotique & Vision
2
What is semantic navigation ?
Using high-level “semantic” information to navigate
ENSTA ParisTech
Thème Robotique & Vision
3
What is a semantic map ?
Non-semantic
ENSTA ParisTech
Semantic
Thème Robotique & Vision
4
Défi Carotte – 2010 - 2012
Robotics competition
– Organized by French Armament Procurement Agency (DGA) and
Research Funding Agency (ANR)
– 5 selected and funded teams
– Artificial environment, 100 m2
Goals
– Autonomous exploration
– Metric mapping
– Semantic information
• Room segmentation
• Ground/Wall types
• Objects type and position
ENSTA ParisTech
Thème Robotique & Vision
5
Video
ENSTA ParisTech
Thème Robotique & Vision
6
Scene segmentation
Original scene
Floor removal
Down-sampling Floor projection
ENSTA ParisTech
Clustering
Thème Robotique & Vision
Walls removal
Interpretation
7
Feature computation
Texture feature
• SIFT, passed through a vocabulary
Color features
SIFT
• Histogram of hue
• Histogram of Normalized
RGB channels
Color
Shape feature
• Surflet-pair relation histogram
Shape
ENSTA ParisTech
Thème Robotique & Vision
8
Object Database
52 objects
6 angles of view
100 snapshots / angle
1 to 4 meters distance
ENSTA ParisTech
Thème Robotique & Vision
9
Classification
Classification
• Concatenate feature vectors
• Feed-forward neural network
Color
Shape
SIFT
Position awareness
• Average scores of superposed
objects in the map
Feed-forward Neural
Network
P
Class
ENSTA ParisTech
Thème Robotique & Vision
10
Resulting semantic map
ENSTA ParisTech
Thème Robotique & Vision
11
Recognition rates
SIFT
ENSTA ParisTech
trgb
Thème Robotique & Vision
hue
SPRH
12
Perspectives
Classification
• Classification depends on success of segmentation
• Neural network is hard to interpret
Database / Hard supervision
Public databases
ENSTA ParisTech
Thème Robotique & Vision
13
Vote-based Local Recognition
ENSTA ParisTech
Thème Robotique & Vision
14
Unsupervised object discovery
– Using laser range finder only
ENSTA ParisTech
Thème Robotique & Vision
15
Example trajectories
Cluster « Sofa »
Cluster «Stool »
ENSTA ParisTech
Thème Robotique & Vision
16
Recognition rates
Halfway around object
ENSTA ParisTech
Thème Robotique & Vision
17
Thank you