Comments on: Introduction to Dynamic Time Warping

By: noname

noname — Sun, 11 Dec 2011 16:56:02 +0000

Hi, an open-source implementation of various dtw algorithms is available for R at http://dtw.r-forge.r-project.org . It has an extensive manual and is callable from Python. Should work well for rapid experimentation.

By: lewis

lewis — Thu, 24 Nov 2011 11:03:09 +0000

Hey man, Im working on a gesture recognition engine using kinect to capture data as my uni dissertation, this was extremely helpful! its the most easy to follow description of DTW i found on the net. quality effot!

By: amira

amira — Sat, 04 Jun 2011 05:40:06 +0000

#of coulmn of y matrix= # of coulmn of x matrix=col=N
#of row of matrix x = row_x , # of row of matrix y =row_y

so the representation look as follow:
A=[x1,...........,x"col";x2,............x''col";................;x"row_x,...........,x"col", y"row_x+1",.......y"col";................;y"row_y".....,y"col"] “is this representation is OK”

all the x cordinate in tthe x first rows followed by y cordinate in next rows.
and matrix y the same.
so iwill use : D[i,j]^2 = (A_xi – B_xj) ^2 + (A_yi – B_yj)^2 ..to replace the line code come belowe
where N =#of col of Matrix A
where M =#of col of Matrix B

for i=2:N
for j=2:M
D(i,j)=d(i,j)+min([D(i-1,j),D(i-1,j-1),D(i,j-1)]);//this line code to replace
end
end

why u calculate D[i,j]^2 not D[i,j] ?
can u help me to reprsent distance matrix u have written using matlab .?

when i plot matrix A and matrix B in matlab give error”that they must have same dimension”
how that if the main advantage of DTW to match vector with different length?

By: Mathieu

Mathieu — Thu, 02 Jun 2011 09:42:56 +0000

Say you have A = ((A_x1,A_y1), …, (A_xn, A_yn)) and B = ((B_x1,B_y1), …, (B_xm, B_ym)), then the cell (i,j) of the distance matrix I mentioned above is D[i,j]^2 = (A_xi – B_xj) ^2 + (A_yi – B_yj)^2.

By: amira

amira — Thu, 02 Jun 2011 00:21:44 +0000

thanks for ur support, i want to ask how to use DTW for matching humana motion trajectories as i represent human action as set of trajectory.

each point in trajectory has (x,y) dimension . really i have two matrix one for x cordinate of all trajectory and second for y cordinate of all trajectory(all trajectory belong to single human action ex:human walk)

so i need two integrate two matrix in single matrix and match resulted matrix refrence matrix with different length

could u help me to modify DTW for human action trajectory matching

By: Mathieu

Mathieu — Mon, 18 Apr 2011 09:33:03 +0000

Once you have computed the matrix, you need to perform backtracking to find the best path (the red dots on the picture above). From this path, you can retrieve the alignment easily. For example, from the picture above, you can see that the point n, n-1 and n-2 are aligned with the point m. Backtracking just consists in applying the recursive rule backward. Google “DTW backtracking” for more information.

By: Shardul

Shardul — Sun, 17 Apr 2011 17:00:23 +0000

Hey Mathieu,

I am trying to use DTW to just align time series data on which I wish to perform K means clustering and subsequently an HMM. But I have not been able to find on how to use DTW just to align two time series data and not use it to classify. Any suggestions?

Thanks,
Shardul

By: Mathieu

Mathieu — Sun, 09 Jan 2011 18:10:18 +0000

Oui j’utilise les methodes a noyau pour l’instant uniquement comme application. J’ai une idee de noyau qui pourrait utiliser une distribution et j’aimerais commencer a m’interesser a MKL egalement.

Je serai a MPS (http://daemon.ice.uec.ac.jp/MPSPortal/events/mps82cfp, Miyazaki, 7 mars) et vraisemblablement a IBIS-ML (http://ibisml.org/ibisml004, Osaka, 28 Mars).

By: marco

marco — Sun, 09 Jan 2011 16:07:59 +0000

Ah bon! si c etait juste un bug pas de probleme!! Je croyais que ca t’avais paru hors-sujet et je me demandais pourquoi…

je vois de quelle presentation tu parles ;) malheureusement les idees que je souleve dans cette video ne sont pas tres utiles en pratique, les quantites qui m interessaient dans ce papier sont incalculables et pas vraiment approximables non plus… (quoique, y a l air d y avoir pas mal d efforts dans ce sens la dernierement). Mais c est vrai que de maniere generale c est important de pouvoir mesurer la similarite entre deux histogrammes. Tu t’interesses aux methodes a noyaux en ce moment? En tous cas felicitations pour ton journal, c’est sympa de lire des articles de vulgarisation / code comme les tiens.

By: marco

marco — Sat, 08 Jan 2011 10:39:18 +0000

un peu surpris de voir que mon precedent commentaire a ete efface et classe comme du spam alors qu’il parlait de travaux directement relies au DTW.