Different Distance in Procrustes

[guisf]

• [ ] In addition of computing norm of the landmark points, after doing Procrustes alignment, generate an image where points inside the contour are filled with 1, remaining points are 0.
• [ ] Cluster these new images with k-means.
• [ ] Plot miss-classification error versus N , and compare with standard k-means.
• [ ] Try simpler shapes, ellipses, rectangles, with noise.

@jovo

[jovo]

not instead, in addition too.

please clarify "more careful analysis of error"

in particular, topological constraint and error vs. n.

On Wed, Oct 5, 2016 at 9:40 AM, guisf notifications@github.com wrote:

-[ ] Instead of computing norm of the landmark points, after doing Procrustes alignment, generate an image where points inside the contour are filled with 1, remaining points are 0. -[ ] Cluster these new images with k-means. -[ ] Do more careful analysis of error, and comparison with standard k-means. -[ ] Try simpler shapes, ellipses, rectangles, with noise,

@jovo https://github.com/jovo

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/neurodata/non-parametric-clustering/issues/12, or mute the thread https://github.com/notifications/unsubscribe-auth/AACjcnSG4XxyBLOFkM3ldbuDJLYAYA2oks5qw6jbgaJpZM4KO0T9 .

the glass is all full: half water, half air. neurodata.io, jovo calendar https://calendar.google.com/calendar/embed?src=joshuav%40gmail.com&ctz=America/New_York

[guisf]

Fixed a bug on the code. Actually the previous shape clustering with K-means is doing reasonable. It usually cluster better than pure k-means. For instance for digits [3,5,7] it clusters with accuracy about 90%, while k-means is 75%. Check: http://nbviewer.jupyter.org/github/neurodata/non-parametric-clustering/blob/master/shape_clustering.ipynb.

[jovo]

that's a start, but for each experiment, repeat 10 times, and for each, plot both the absolute error, and the difference of errors (so positive means procrustes does better)

and then add euclidean \circ procrustes too.

On Wed, Oct 5, 2016 at 5:59 PM, guisf notifications@github.com wrote:

Fixed a bug on the code. Actually the previous shape clustering with K-means is doing reasonable. It usually cluster better than pure k-means. For instance for digits [3,5,7] it clusters with accuracy about 90%, while k-means is 75%. Check: http://nbviewer.jupyter.org/github/neurodata/non- parametric-clustering/blob/master/shape_clustering.ipynb.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/neurodata/non-parametric-clustering/issues/12#issuecomment-251812154, or mute the thread https://github.com/notifications/unsubscribe-auth/AACjcoTkJHAT-z22MsjlpEkD_CJ2fsqAks5qxB3LgaJpZM4KO0T9 .

the glass is all full: half water, half air. neurodata.io, jovo calendar https://calendar.google.com/calendar/embed?src=joshuav%40gmail.com&ctz=America/New_York