extracting adaptive contextual cues from unlabeled regions
DESCRIPTION
Extracting Adaptive Contextual Cues From Unlabeled Regions. Congcong Li + , Devi Parikh * , Tsuhan Chen + + Cornell University * Toyota Technological Institute at Chicago. Object Detection with Context. Previous: Focus on labeled objects but neglect unlabeled regions . plant. plant. - PowerPoint PPT PresentationTRANSCRIPT
Extracting Adaptive Contextual Cues From Unlabeled Regions
Congcong Li+, Devi Parikh*, Tsuhan Chen+
+ Cornell University* Toyota Technological Institute at Chicago
International Conference on Computer Vision 2011Poster ID:
2-46
Object Detection with Context
plantchair
sofa
plant
Previous: Focus on labeled objects but neglect unlabeled regions
Labeled vs Unlabeled
55% 45%72%28%
MSRC dataset PASCAL 07 dataset
Human Study: unlabeled regions help
Is unlabeled region useful?
Our View: Leverage unlabeled regions
plant
‘plant’ context
Our view: Extract adaptive context
Inter-object Intra-objectScene
• Ours: Context at adaptive granularities Multi-level Interactions!
• Prior works: Context at fixed granularity
20%
EXO: expand fixed ratio Scene: whole image
Contextual-Meta Objects (CMO)
Algorithm: discovering contextual regions
Database
Extent-based Clustering
Content-based Clustering
... …
Learn “object” Models... …
Context Detector
Results on PASCAL 2007
31.5
32
32.5
33
33.5
34
34.5
Scene EXO CMO
Adaptive granularity helps!
32.5
33
33.5
34
34.5
35
35.5
Labeled Unlabeled Combined
Unlabeled: complementary context
fixedgranularity
adaptive
Results: improve multiple detectors!
22
23
24
25
26
27
28
24
26.8
Labeled objects Unlabeled regions
Dalal-Triggs10
11
12
13
14
15
16
12.6
15.6
Implicit Shape Model30
31
32
33
34
35
36
32.3
34.2
Part-based Model
Can employ any object detector to learn the contextual “object”!
Results: provide spatial prior for OOI
Contributions
• Extracting contextual cues from unlabeled regions
• Capturing contextual interactions at varying levels: Scene, Inter-object, Intra-object
• Extracting contextual regions by learning “object” models using any object detector
• Intelligently leveraging existing techniques: easily accessible to community
Thank you!
International Conference on Computer Vision 2011Poster ID:
2-46
Visit our project page: http://chenlab.ece.cornell.edu/projects/AdaptiveContext/