Weekly meeting 7 Dec 2016: Luis and Tom
==========================
What I have done:
===========
– I showed Tom the results without data augmentation and just training with one of the three training videos
(there are 4 available for training, but I used one for validation of the CLR). The total Dice for the dataset was
around 64%. This low number could be due to two factors: 1) I did not use the validation set for training and
2) data augmentation was not used.
Comments by Tom:
============
– Do not spend more time on CLR optimisation, as it is now is ok.
– Implement Dice loss layer for journal paper.
– Test inference speed of FCN-32s vs FCN-8s.
– Implement ESM optical flow framework.
– Add the previous image of the video to the FCN input so that motion is accounted as input to infer the segmentation mask.
– Try the previous point with transfer learning from DAVIS dataset as this is video segmentation with good annotations.