Skip to content

Latest commit

 

History

History
88 lines (72 loc) · 3.98 KB

README.md

File metadata and controls

88 lines (72 loc) · 3.98 KB

Delay Embedding

Time series modeling and classification based on delay embedding. This is an improved version of "Derivative Delay Embedding: Online Modeling of Streaming Time Series".

Pre-requisites

Matlab (the code has been tested on Matlab 2015a)

Running the test

Run MAIN.m

>> MAIN

Preliminary Results

The running print on MSR Action 3D dataset is shown as follow

Processing the MSR_Action3D dataset 
Trained 50 / 284
Trained 100 / 284
Trained 150 / 284
Trained 200 / 284
Trained 250 / 284
tested 50 / 273
tested 100 / 273
tested 150 / 273
tested 200 / 273
tested 250 / 273
Training time: 2.670sec, 0.009sec per sample
Testing time: 16.898sec, 0.062sec per sample
Accuracy = 93.77%

Compare to the state-of-the-art algorithms

Moving poselets
(ICCV2015)		 dRNN
(ICCV2015)		 HBRNN
(CVPR2015)		 Actionlets & Poselets
(CVPR2016)		 Our method
(2016)
Accuracy 93.6% 92.03% 94.49%* 93.0% 93.77% / 94.52%*

Note: the * marker denotes the results from subsets, which is usually higher than that from the whole dataset

Folders and Files

  • DE implements Delay Embedding
  • delayEmbeding.m implements 1-D delay embedding
  • delayEmbedingND.m implements multi-dimensional delay embedding
  • MGM learns Markov Geographic Model
  • createGrid.m creates discretized embedding space.
  • add2Trans.m records learned transition
  • Trans_Prob.m computes transition probability
  • HDist.m calculates distance between a testing sample and learned model (transition probability)
  • data
  • MSR_Action3D.mat is the MSR Action3D dataset
  • UCI_CharacterTrajectories.mat is the Character Trajectories Data Set from UCI
  • setting_MSR.m and setting_UCI.m are settings for the two datasets used in MAIN.m
  • utilities
  • confusionMatrix.m plots the confusion matrix
  • defaultColors.mat stores the default color map of Matlab
  • lowpassFilter.m performs low-pass filter to filter the raw data

Related Work

Zhifei Zhang, Yang Song, Wei Wang, and Hairong Qi. "Derivative Delay Embedding: Online Modeling of Streaming Time Series". The 25th ACM International Conference on Information and Knowledge Management (CIKM), 2016. [PDF]

@inproceedings{zhang2016derivative,
  title={Derivative Delay Embedding: Online Modeling of Streaming Time Series},
  author={Zhang, Zhifei and Song, Yang and Wang, Wei and Qi, Hairong},
  booktitle={Proceedings of the 25th ACM International on Conference on Information and Knowledge Management},
  pages={969--978},
  year={2016},
  organization={ACM}
}

Presentation on CIKM2016. If it is not availabel, please refer to the PDF version.