Hi Tom,
Meeting minutes for the weekly meeting: March 8, 2017.
What I have done:
===========
– Refactoring the IROS Python code. I realised that V-Net Dice loss layer is binarising the predicted
probabilities before calculating the loss.
Discussion:
=======
– Boundary improvement options: loss functions including distance to the border or ground truth + labels on the edges.
– A possible motivation for the use of motion/temporal information is to regularise the regmentations of unseen instruments.
TODO list:
=========
– Prepare 5 min slides for the upgrade viva (already done!).
– Try Grabcut from OpenCV on the predicted labels of ToolNetH (low priority).
– Submit IROS support material (high priority).
– Implement multi-scale boundary loss (low priority).
– Ask Max about the TMI (already done!): he has no idea why it is going so slow.
– Deadline for the journal: 31st March (high priority), draft one week before.
Cheers,
Luis.