Dan Munoz
Willow GarageSummer Finale Presentation
Mentor: Kurt Konolige
CMU Research
Building
Vegetation
GroundTree trunk
Sky
Support
Vertical
Geometry Surface Estimation (Hoiem et al.)
3-D Point Cloud Classification
2
Contextual Classification with Functional Max-Margin Markov Networks D. Munoz, J. A. Bagnell, N. Vandapel, M. Hebert
CVPR 2009
Outdoor Classification
3
Vegetation Tree-trunk
Ground Building
Summer 09
4
Floor Ceiling Column
Chair Table Wall
Challenges Training data
Discriminative features
Learning
5
Training data Issues with getting 3-D labels from 2-D images
Manually created labeled dataset of room-sized objects (chairs, tables, trash cans, etc.) in PCD format 6
Discriminative Features descriptors_3d
• Utilizes PCML routines
• Similar interface with image descriptors_2d (Alex T.)
• Parallelized with OpenMP*
Example features:
Works over groups/clusters of points as well• point_cloud_clustering
(k-means, single-link, local neighbors)7
Image from Johnson and Hebert 1999
Learning functional_m3n
• ROS agnostic
• Can do online learning
• Extended implementation of CVPR’09 work
1 min tutorial…
8
Independent Classification
9
Local Interactions
10
Using Higher Order Information
11
Colored by elevation
Region Based Model
12
Simple Algorithm For T iterations
• Classify with current model
• Create training set D from misclassifications (Over features from each clique)
• Train favorite classifier OpenCV regression trees
• Add classifier to model
D
13
Experiments
14
Table-top Objects
15
Table-top Objects
16
Only 3-D features (worst example)
Stapler Mouse
Mug Background
Adding Image Features
17
Results
18
Stapler Mouse
Mug Background
Results
19
Stapler Mouse
Mug Background
Results
20
Stapler Mouse
Mug Background
Results
21
Stapler Mouse
Mug Background
Results
22
Stapler Mouse
Mug Background
Results
23
Stapler Mouse
Mug Background
Room-sized Objects
24
Floor Ceiling Column
Chair Table Wall
Cabinets
Room-sized Objects
25
Floor Ceiling Column
Chair Table Wall
Cabinets
Room-sized Objects
26
Floor Ceiling Column
Chair Table Wall
Cabinets
More experiments required There were no cabinets off the ground in training set
27
Floor Ceiling Column
Chair Table Wall
Cabinets
Available Future Work descriptors_3d
• Faster neighborhood data structure (Marius, Ethan)
• Point Histogram Features (Radu, Gary)
• 3-D Chamfer distances (Marius)
point_cloud_clustering• Ground plane removal and then clustering 2-D projections
(Caroline)
• Mean-shift, etc.
functional_m3n• Random Forests, Neural nets (OpenCV)
• Boosted spheres (Alex T.)
28