Author: luisherrera

Meeting minutes of the weekly meeting with Tom on the 26 Oct 2016: What I am doing: ========== Replicating the MICCAI results to send them to Max for the TMI journal paper. RSG meeting: ======== We will organise a meeting

Meeting minutes for 31 Aug 2016: Tom and Luis ===================== What I have been doing: – I have been working on the optimisation of the CLR boundaries. Particularly on the optimisation of the base_lr and max_lr.    Now the train

Things I have done: – Refactor the code of the cnn_test. Solve problems with using gift-grab dynamically linked, embedded in the code now). – Still solving synchronisation bugs. Ideas from Tom for creating a discriminative and data-driven feature descriptor: –

Present in the meeting: Luis and Tom. – I showed Tom the distance maps to the mean RCD of the background Gaussian. They seem to signal properly where the tool is in the image so this information can be used

Present in the meeting: Luis and Tom. The histograms show that the RGB of the background is Gaussian. The foreground is not, it looks more like a Gamma. TODO for next week: – Remove burnt pixels and super black pixels

Present in the meeting: Luis and Tom. – I showed Tom the results of the clustering of superpixels using a covariance matrix to describe each superpixel and RGB to describe the individual pixels inside a superpixel. The superpixels where obtained

Hi Tom, What I showed: – Today I showed to Tom the segmented images with the superpixel-based regularisation and markers on the borders between the superpixels. I changed the sigma parameter. It seems that the small superpixels were due to

– I showed the Dice coefficients for the runs with GMM and SMM. GMM gets better Dice coefficients and on top of that RGB gives better results than all the other feature vectors tested. From the runs that I did

Ideas: – Use super pixel as a regularisation strategy by using voting of the probabilities to decide if the super pixel is tool or not. Combine this approach with clustering the super pixels by using RCD and GMM or SMM.