2011_08_05

Post date: Aug 6, 2011 11:14:02 PM

This is one productive day. Last night I ran the ITSBN version 3.0.2 on the Berkeley dataset and got the pretty good results. So I added some of them to the MLSP2011 paper, changed some texts, updated the results, got it submitted already.

Also I successfully dusted off my unsupervised boundary evaluation code, but it still needs some more work before releasing to public. Anyway, ITSBN competitively beats supGMM by 0.004 on the F-measure. This makes me think that "perhaps superpixel makes structured inference less important." Remember when comparing TSBN with GMM, then former significantly beats the latter, but not significantly when replacing pixel by superpixel.

On my ONR work, my DT code seems to work on the 7-feature data. I'm not so sure about the performance though-- let's see.