What I have done:
===========
* I showed the complete draft of the upgrade report.
* I showed that the segmentation results (evaluated as the Dice coefficient of the foreground) on the MICCAI robotic dataset with
and without data augmentation (flipping the images around the vertical axis) seem to indicate that the data augmentation is helpful.
More tests with the other two datasets are needed.
* I showed the results comparing the segmentation results when training with a CLR and when training with a fixed learning rate.
At least for the MICCAI robotic dataset, the results seem to indicate that a ~3% imporvement on Dice coefficient ios achieved by using
and optimising the parameters of the CLR.
What Tom said:
==========
– Change the name of the assistive technologies to Active Tracking Systems.
– Change the first section of the literarure review to “Detection, segmentation and tracking in Computer Vision”.
– Change the title of the section about adhoc methods to “ad-hoc methods specialised for surgical vision”.
– Finish the literature review by saying that CNNs couple feature extractiong with classification in a single framework.
– Move all the methods of the literature to the first section and delete the section called “Methods”.
– Make the subsection called “Specialised…” an actual section and not just a subsection.
– Future plans: a sensible plan would be to concentrate on improving the segmentation of surgical tools, mainly with
the focus put on two aspects:
1) Incorporate temporal information (motion) into the current framework. This means for example testing transfer
learning with the DAVIS dataset.
2) Network compression with the idea of maintaining the representative power but reducing the size of the network
so that the inference takes less time, allowing then for other methodological improvements such as the use of more
than one network.