Abnormal Object Detection by Canonical Scene-based Contextual ...
Sangdon Park 2012.10.15. Abnormal Object Detection by Canonical Scene-based Contextual M odel Introduction Problem Statement Abnormal Object Detection (AOD) Input Output Which objects are abnorma l?
2 Introduction Problem Statement Three types Co-occurrenceviolating abnormal object of Abnormal Objects Position-violating abnormal object Scale-violating abnormal object 3
Introduction Motivation Increasing number of Abnormal Images Photosho p Artist Applicable to Visual Surveillance Duck Climbing 4 Introduction Motivation Limitation of the conventional method(1)
NOT affluent object relations Tree-relation among objects quantitative object relations affluent context types prior-free object search (1) M. J. Choi, A. Torralba, and A. S. Willsky, Context Models and Out-of-context Objects, To appear in Pattern Recognition Letters, 2012. 5 Introduction
Contributions Solve new emerging problem Abnormal Object Detection Novel latent Model Generative model for AOD Satisfies four conditions for AOD Especially, affluent object relationships to strictly handle geometric context New abnormal dataset
object-level annotation 6 Agenda Conventional Method Proposed Method Evaluations 7 Conventional Method Tree-based model Tree-based Co-occurrence model Tree-based
support model Efficient, but lack of relationship among object 8 Proposed Method Overall process 9 Proposed Method Image representation Object-level image representation Represent image by a set of bounding boxes that are
extracted by object detectors Each image consists of bounding boxes (=100, in this paper) Undo projectivity Transform image coordinate to camera coordinate by simple triangulation Represent position and scale information altogether 10 Proposed Method
Main Idea Identify abnormal ones! Which object is abnormal? Which object is less co-occur, floated/ sunken, or big/small? Define dist. of normal data & Compare? Compare the input with the distribution of normal objects Check likelihood of input given the dist. How to represent the distribution of normal scene? Construct the Canonical Scene (CS) model
How to compare the input scene with the normal scene? Matching transformation T for CS Similarity measure to compare the input scene and transformed CS 11 Proposed Method Model Define Canonical Scene Outdoor CS
Natural distributions of normal objects Less co-occurring objects does not exist Objects are on the ground plane Follows leaned truncated Gaussian distribution 12 Proposed Method Model Define
Matching transformation matching transformation & similarity measure T: 2D isometric transformation Similarity measure msT,l ( Lo ,n , xo ,n ) p( K o , n , xo , n | s, lo ,n , T ) o ,n 13 Proposed Method Model Return to the goal
Mod el Decompo se p ( s, l, c, T ) p ( y | c) p(K , x | s, l, T )dK Prior model Appearance Model Location(Contextual) Model Prior on latent variables
Defined by previous similarity measure Defined as conventiona l model 14 Proposed Method Model Generative model Isometry Parameters of
Canonical Scene 15 Proposed Method Inference by Pop-MCMC Advantages of Pop-MCMC Multiple Markov chains with genetic operations escape from local optimum Efficient when the objective function is multimodal and/or high dimensional
16 Proposed Method Learning Learning strategy Estimate T, thus making complete data Assumes all objects in normal images are on the gro und plane T is a transformation that transform ground plane in w orld coord. to slanted plane in camera coord. T 1
Algorithm 17 Evaluation New Abnormal Dataset Only abnormal objects are annotated Scene types are also annotated #images
Evaluation Quantitative comparisons Proposed method(red) outperforms the baseline(green ) CO+SUP: M. J. Choi, A. Torralba, and A. S. Willsky, Context Models and Out-of-context Objects, To appear in Pattern Recognition Letters, 2012. 19 Evaluation
Qualitative comparisons Because of affluent object relation, floating person is detected as most abnormal objects 20 Evaluation Qualitative results Only top-5 most abnormal objects are represented
21 Conclusion Novel Model for Abnormal Object Detection New abnormal dataset Generative model Satisfies four conditions for AOD Especially, affluent object relationships to strictly handle geometric context State-of-the-art performance Limitations
Learning Full parameter learning is required Annotation errors Cannot estimate ground plan e strictly poor performance on detecting scaleviolating abnormal objects 22
Isaiah 40:31 - But those who wait on the Lord Shall renew their strength; They shall mount up with wings like eagles, They shall run and not be weary, They shall walk and not faint. The Lord can help you...
Logical operations. Arithmetic operations. Boolean Algebra. Algebraic representation of logic. ... Division. Unsigned addition. Suppose we have a computer with 4-bit words. What is the unsigned value of 7 + 7? 0111 + 0111. What about 9 + 9? 1001...
LAMI'S THEOREM α Fig (2) F1 F3 F2 Space Diagram(SPD):The sketch showing the physical conditions of the problem, like, the nature of supports provided; size, shape and location of various bodies; forces applied on the bodies, etc., is known as...
Registry includes: Information about 29 print archive programs or collections, 55,000 bibliographic records and associated holdings from 12 programs/collections and 66 institutions, ... UCL Shared Print. 2,456. 4. Western Regional Storage Trust. WEST. 10,845. 28. Print Archiving by Program.
Franklin D. Roosevelt: Quarantine the Aggressors Chicago, October 5, 1937. World War II. The Road to War. Kellogg-Briand Pact 1928. After breaking the isolationist policy during World War I in 1917, the United States returned to their reclusive ways during...
Limited visibility and clarity of downhole toolstring actions and status . Limited depth precision. Limited downhole functionality . Over-dependence on slickline operator's experience level. 5/20/2015. Historical Slickline Perforating. No real time correlation. Parameter run.