10472 10316 Mentor: Prof.Amitbha Mukerjee amit@cse.iitk.ac.in 4 - - PowerPoint PPT Presentation

▶
10472 10316
SMART_READER_LITE
LIVE PREVIEW

10472 10316 Mentor: Prof.Amitbha Mukerjee amit@cse.iitk.ac.in 4 - - PowerPoint PPT Presentation

Nov 13, 2022 680 likes •876 views

P.Yaswanth Kumar Jitendra Kumar 10472 10316 Mentor: Prof.Amitbha Mukerjee amit@cse.iitk.ac.in 4 tasks 4 tasks Counting number of 1) characters A 2) green bars 3) horizontal bars 4) vertical bars 4 tasks Counting

slide-1
SLIDE 1

P.Yaswanth Kumar Jitendra Kumar 10472 10316 Mentor: Prof.Amitbha Mukerjee amit@cse.iitk.ac.in

slide-2
SLIDE 2

 4 tasks

slide-3
SLIDE 3

 4 tasks  Counting number of

1) characters ‘A’ 2) green bars 3) horizontal bars 4) vertical bars

slide-4
SLIDE 4

 4 tasks  Counting number of

1) characters ‘A’ 2) green bars 3) horizontal bars 4) vertical bars Training HMM’s for each task

slide-5
SLIDE 5

 4 tasks  Counting number of

1) characters ‘A’ 2) green bars 3) horizontal bars 4) vertical bars Training HMM’s for each task Task inference for a new given Eye gaze trajectory

slide-6
SLIDE 6

 Yarbus Process

[2] [2] Task

slide-7
SLIDE 7

 Yarbus Process

[2] [2]

Many methods exist for Yarbus Process.

Task

slide-8
SLIDE 8

 Yarbus Process

[2] [2]

Many methods exist for Yarbus Process. Inverse Yarbus Process ?

Task

slide-9
SLIDE 9

 Inverse Yarbus Process :

TASK ?

slide-10
SLIDE 10

 For each task :

Get the people Give them task Collect Eye Gaze Trajectory obtained.

slide-11
SLIDE 11

 For each task :

Get the people Give them task Collect Eye Gaze Trajectory obtained. Count no of A’s

slide-12
SLIDE 12

 For each task :

Get the people Give them task Collect Eye Gaze Trajectory obtained. Count no of A’s

Fixation points

slide-13
SLIDE 13

Observed Sequence of states obtained from the above trajectory is 9,3,4,5,6,14,15,15,9,18,22,23,24,25,26,27,28,29,29,31,32,33,3 3,35,35

slide-14
SLIDE 14

 BAUM-WELCH ALGORITHM:

Constructs a HMM for each task by taking the

  • bserved sequence of states matrix obtained.

λ = (A,B, π) A = State Transition Matrix B = Observation Probability Matrix Π = Initial State Observation Matrix [ π, A, B ] = dhmm_em(data, πe, Ae, Be ,’max_iter’,5);

slide-15
SLIDE 15

 FORWARD ALGORITHM :

For a new given observation sequence, find the likelihood of each task using their HMMs Loglik = dhmm_logprob(data_new, π, A, B); Task with maximum loglikehood value is the REQUIRED TASK.

slide-16
SLIDE 16

For 8 test data sets loglikelihood values obtained are :

slide-17
SLIDE 17

[1] Haji-Abolhassani, A. and Clark, J.J., "Visual Task Inference Using Hidden Markov Models", proceedings of International Joint Conference on Artificial Intelligence (IJCAI), pp. 1678--1683, 2011 [2] A.L. Yarbus. Eye movements during perception of complex objects. Eye movements and vision, 7:171–196, 1967. [3] Source Code: http://www.cs.ubc.ca/~murphyk/Software/HMM/hmm.html

slide-18
SLIDE 18

THANK YOU !! QUESTIONS ?