close

Вход

Забыли?

вход по аккаунту

код для вставкиСкачать
Week 8
Emily Hand
UNR
Sub SVM

Guang Implemented a function blockSVM

Takes all blocks from a template

Determines confidence of each block



With respect to the template's SVM model
Returns a confidence map
Tells us which parts of a template are
contributing the most to the confidence of the
SVM
Framework Idea

Assumption: After selected, the person does not
jump behind an object



Safe to assume
High SVM confidence for whole template

We have 1 SVM for the whole template

Don't test individual blocks
If confidence is low

Test individual blocks and display occlusion
Update Strategies

Blended Template

Threshold Retrain


Only retrain if the number of confident blocks is
above a threshold
Find a template for regions that are not
occluded and test with this

Introduce a 2nd SVM
Blended Template



SVM for entire template is used to find the
location in the next frame
Confidence of each block extracted using
blockSVM function
A new template is constructed from positive
blocks

Blended with most recent positive blocks to form a
complete template
Blended Template

Good results so far...

Will run more sequences today.
Threshold Retrain



SVM for entire template is used to find location
in the next frame
Confidence of each block extracted
If a large number (say 60%) of the blocks have
a high confidence, then retrain the entire block.

Accept small changes

Don't damage the SVM model too much
Occlusion Template

SVM for entire template

High Confidence

Continue Normally


Retrain Classifier
Low Confidence

Find confidence of individual blocks in template

Use those to train a new SVM model

Scan with template and extract the confidence
blocks and test with the partial SVM model
Occlusion Template

When searching for
the template in the
neighborhood, only
these blocks will be
extracted from each
sample and tested
in the partial SVM
model.
Occlusion Template

Yesterday it was very slow

I was able to speed it up quite a bit

Not getting the results we would like

Found the bug, but need to fix it.

Run lots of sequences.
Thresholding

No fixed threshold



Range based off of
previous confidences
If confidence is within
this range, then it is
most likely the object
of interest
Testing idea today
1/--страниц
Пожаловаться на содержимое документа