Final

Question 1

Challenges to Trackin

Answer 1

1) Many Places its hard to compute optical flow
2) There can be large displacements since it could be moving rapidly. You need to take the dynamics into account.
3) Errors would compound or drift
4) Occlusions, dis-occlusions

Question 2

Shi Tomasi Feature Tracker

Answer 2

Only compute motion where you should.

1) From frame to frame, track with lucas kanade and pure translation model
2) For each of the points you have tracked, is there an affine model to represent this translation

Question 3

Tracking with Dynamics

Answer 3

This is used if the displacements are large because there is fast movement. Given a model of expected motion, preditucs where the object will occur in the next frame, even before seeing the image.
restricuts search of the object
imrpved estimates

Question 4

Detection vs. Tracking

Answer 4

the incormproiation of dynamics is the differece between tracking and just detection. The detection is independent each time. Tracking we predice with estiamted dynamics and then update based on upon measurements.

Question 5

Assumptions in Tracking

Answer 5

Continous Models of motion Paramenter
Objects do not appear and disappear instantly
Camera is not moving instantly
Gradual change in pose between scene and object

Question 6

Prediction

Answer 6

What is the next state of the object given past measurements

Question 7

Correction

Answer 7

Compute an update destiamte of the state from preduction and new measurments

Question 8

Tracking

Answer 8

The process of propagation this posterior distribution of state given measurements across time

Question 9

Tracking Simplyfying Assumptions

Answer 9

Only the Immediate Past maters

Measurements depend only on the current state

Question 10

Tracking as Induction

Answer 10

Base Case:
Assume an initial prior tha tpredicts in the absense of any evidence
At the first frame tak ea mesurment and correct a predocution
Preditct for frame t + 1
Correct for frame t + 1

Question 11

Prediction

Answer 11

Given t-1 information, then make a guess of t

Question 12

Correction

Answer 12

now predict again with the yt using again. This ues Bayes rule P(A|B) = P(B|A)P(A)/P(B)

Question 13

Kalman Filter

Answer 13

Linear Dynamics Models in gaussian noise

State densities are gaussian

Question 14

Partical Filterting

Answer 14

Relies on non gaussian distributions

Question 15

Perturbation

Answer 15

This is equivalent to the dynamics model

Known ways the world should change that are modeled by the filter

Question 16

Bayes Filter Framework

Answer 16

1) Priort probability of the system state p(x)
2) Action (dynamical system) model)
p(x..t|u..t-1, x..t-1)
3) Sensor Model (likelihood) p(z|x)
if the object is someplace what is the likeliehood of my measurement
4) Stream of Observations z and action data
u: data = {u1,z2,….

Question 17

Object Categorization

Answer 17

Given some number training instances find images with recognized labels from the training instances

Question 18

Risk

Answer 18

is the sum of the loss function times the probability of that incorrect classification happening