Description
An attempt to use Support Vector Machine (SVM) to quantify the weight of each metric in determining the likelihood of frames being adjacent was done with mixed success. Originally a classifier was trained to determine if frames were adjacent but due to the highly skewed set of training examples (only two adjacent frames out of a set of n examples) the classifier had poor performance.
Train an SVM classifier to determine if frames are adjacent
steps of the Code :
Detect features in video frames
Calculate match metrics between all frames
Calculate metrics from matched features
Calculate classes for adjacent frames
Restructure data for classification training
Train SVM to classify adjacent frames
Check model
Classify test data and evaluate performance
Save SVM model
Use SVM classifier to detect adjacent frames
Order by algorithm :
1) Pick random starting frame
2) Find nearest neighbor (top percentage of matches then pick min distance)
3) Check if closer to start or end of constructed sequence
4) If closer to the end append the frame
5) Otherwise start the search from the start of the sequence and build backwards
6) Continue until all frames sorted
Reviews
There are no reviews yet.