Action Recognition

Download Report

Transcript Action Recognition

Action Recognition in
Temporally Untrimmed Videos
Fatemeh Yazdiananari
Temporally Clipped v.s
Unclipped
• Temporally Clipped: Videos only contain
the action.
• Temporally Unclipped: Videos contain both
the action and non-action.
• Temporally Unclipped is a real-world
representation of videos. Action recognition
needs to be adapted for it.
2
Unclipped Videos
• Contains more then the action
• Determine the temporal location and the
action itself
• Make temporally clipped recognition
methods suitable for unclipped data
• We are considering 4 different methods
The 4 Methods
•
•
•
•
1. Dividing a video into clips
2. Overlapping Sliding Windows in time
3. Spatiotemporal Segmentation
4. Graphical Model: Capturing the
relationship of clips
Baseline Action Recognition
• Using DTF features
– HOG, HOF, MBH, Trajectory
• Bag of Words model
– Feature Vector: each video is represented by a
histogram of visual words
– SVM is used as the classifier
Preliminary Steps
• Download UCF101, DTF, three split files
• Run and understand demos of SVM
• Work on UCF101 baseline
– Write code to load Features, Labels, and
Names of each video.
SVM demos
Ground truth of all data both test and training data
SVM demos
Small unfilled circles are the trained data, filled circles are the
tested data. Only were classified as positive.
8
Code
• Feature matrix (DTF) : (13320, 16000)
• Label Vector : (13320,1)
• Name Vector : (13320,1)
• Next step is to optimize this into a structure for each
video with feature, label, name and index
• Optimization will help me run a comparison with the
Train/Test splits and implement MultiClass SVM
• Next week I will be able to run baseline and get
accuracy percentage of UCF101