Title: Description Logic for VisionBased
1- Description Logic for Vision-Based
- Intersection Interpretation
Britta Hummel
2Motivation
Road Recognition The Model-based Approach
1. Project
Low-dim. geometry model (clothoid, )
2. Compare
3. Update Parameters
- Solved for highly constrained domains (highways)
3Motivation
Intersection Recognition HeimesNagel02
1. Project
2. Compare
3. Update Parameters
- How can we generalize to arbitrary intersections?
4Motivation
Challenges
- High-dimensional
- hypothesis space
- 2. Few features
- - Narrow field of view
- - Massive occlusions
- - Omitted features
- Presence of noise
- - Unmodelled objects
- - Bad feature quality
- Model-based approach becomes ill-posed!
5Motivation
So what now?
- More top-down information flow
- ? start higher up use conceptual knowledge!
- ? move further down parameterize feature
detectors! - Collective classification
- ? simultaneously reconstruct geometry and
semantics!
- Narrow down hypothesis space!
- FOL Representation and FOL Reasoning!
6This Talk
- Motivation
- Architecture
- DL Road Network KB
- DL Inference for Scene Interpretation
- Application
- Evaluation
7Architecture
Enhance Model-Based Vision by Logic
Feature detectors, other KBs,
Project
Update Pars
8This Talk
- Motivation
- Architecture
- DL Road Network KB
- DL Inference for Scene Interpretation
- Application
- Evaluation
9Model of Geometry
Geometric Primitives
GP1
GP3
GP2
Spatial Relations
10Symbol Grounding
Geometric Primitives
Spatial Relations
11TBox
Geometric Constraints
12TBox
Constraints wrt Road Building Regulations
13ABox
Sensor Data Integration
- Partial observability
- ? OWA
- Structurally differing sensor data (e.g. from
map, video) - Distributed sensor data
- Non-UNA identification reasoning
- Open/Closed Domain Data
- (Nominals) / Closed domain assumption
-
- Conflicting/Uncertain Data
- ? BLPs/MLNs/
14This Talk
- Motivation
- Architecture
- DL Road Network KB
- DL Inference for Scene Interpretation
- Application
- Evaluation
15Inference Example I
(Collective) Classification is Abox realization
l21
l22
l11
l12
l13
16Inference Example I
(Collective) Classification is Abox realization
l21
l22
l11
l12
l13
16
17Inference Example I
(Collective) Classification is Abox realization
l21
l22
tr-l11-l21
l11
l12
l13
17
18Inference Example I
(Collective) Classification is Abox realization
l21
l22
tr-l11-l21
l11
l12
l13
19Inference Example I
Link Prediction is Instance Checking
19
20Inference Example II
Link Prediction is Instance Checking
l21
l22
tr-l11-l21
l11
l12
l13
21Inference Example III
Data Association is Identification Reasoning
Positioning Device Map Matching
Video
Digital Map
?
22Inference Example IV
Hypothesis Generation is ?
- Classical logical inference is deductive
- Bio./Mach. Vision is not deductive lots of
hypothetical reasoning, jumping to conclusions,
backtracking if wrong - ? Non-deductive / non-monotonic reasoning
needed! - Abduction
- Poole, Shanahan, Möller
- Introducing procedurality
- NeumannMöller06
- Model construction by transformation into
Constraint Satisfaction Pr. - ReiterMackworth87
- Model construction under Answer Set Semantics
- We have started
23Beautiful Analogies
24This Talk
- Motivation
- Architecture
- DL Road Network KB
- DL Inference for Scene Interpretation
- Application
- Evaluation
25Application
Geometry model generated from DL ground truth
ABox
26Application
27This Talk
- Motivation
- Architecture
- DL Road Network KB
- DL Inference for Scene Interpretation
- Application
- Evaluation
28Summary
- Road Recognition ? Intersection Interpretation
- escape from toy world ? narrow down hypothesis
space - not only bottom-up but also top-down
reasoning - collective classification
- Enhance model-based vision by logical reasoning
- Expressive geometry model
- Generate generic geometric model out of logical
configuration model - Generate and constrain logical model through
logical reasoning
29Evaluation
- Vision
- Sets of knowledge engineers codingmaintaining
large, distributed, modular, semantically
unambiguous KBs for SI - DL
- Wish List ?
- Foundational ontologies / best practices for KB
design for SI - Faster Abox reasoning (gt10 individuals
prohibitively slow on our KB) - Language expressiveness
- Spatial Relations JEPD condition
- Feature chains
- Nominals
- Nonmonotonic reasoning
30Outlook
- Nonmonotonic reasoning with ASP
- Incremental hypothesize test
- Integration with Irina Lulchevas MLN-based
traffic participant classificator - Rule Learning from Training Data
31Thanks
32Description Logic
- Decidable subset of 1st order logic
- Syntax
- Semantics Set-theoretic
33Description Logic
- Axioms form sentences
- A DL Knowledge Base consists of
- Tbox Set of terminological axioms
- ? general domain knowledge
- Abox Set of assertional axioms
- ? knowledge about particular situation
- ( Rulebox )
-
34Description Logic
-
- Classical DL inference services
-
- Non-classical inference
- .