10
Emily Hand UNR Week 10

Week 10 Emily Hand UNR

Embed Size (px)

DESCRIPTION

Online Learning Tracking by detection First Frame Each Following Frame Manually select target Extract features from target Train an SVM Each Following Frame Detect target using SVM model Retrain SVM with old samples and new sample

Citation preview

Page 1: Week 10 Emily Hand UNR

Emily HandUNR

Week 10

Page 2: Week 10 Emily Hand UNR

Online Learning

Tracking by detection First Frame

Manually select target Extract features from target Train an SVM

Each Following Frame Detect target using SVM model Retrain SVM with old samples and new sample

Page 3: Week 10 Emily Hand UNR

Online Learning

Features Histogram of Oriented Gradients (HOG) Local Binary Pattern (LBP)

Both are computed over all 3 color channels 3D Color Histogram

Sliding Window method for detection Scan the template over a small neighborhood

around the previous position of the target. Test each of these templates against the SVM model

Page 4: Week 10 Emily Hand UNR

Occlusion Handling

Partial Occlusion Break up template into small blocks. Determine how these parts contribute to the SVM

detection score Use this information to determine occluded parts of

a template Full Occlusion

Motion Model used for predicted location

Page 5: Week 10 Emily Hand UNR

Occlusion Handling

Other methods: Retrain the classifier even if the target is partially

occluded Damages the SVM model

Blended Template Keep track of all the blocks contributing the most to

the SVM score. Previous Template + Present Template = Blended

Template Retrain Classifier

Page 6: Week 10 Emily Hand UNR

Results

Page 7: Week 10 Emily Hand UNR

Results

Page 8: Week 10 Emily Hand UNR

Results

Page 9: Week 10 Emily Hand UNR

Results

Page 10: Week 10 Emily Hand UNR

Current Work

Partial SVM Find parts that are not occluded Train a new SVM with only those parts and search

for that in the next frames Latent SVM

Each target is made up of 6 parts All 6 parts have their own SVM

Train on first frame Search for each individual part Parts are free to move around.