ToDo-List-Depth

This is the short and long term to-do list of training Network to predect depth

About the training dataset

Enlarge the Dataset

Data generation (down sample)

Fill the holes (pixels that the depth is not available)

About the network

Architecture:

Drop Out (to prevent overfitting)

Residual net (good for deeper net)

Format of Depthmap and loss function

log depth

invers depth

scale invariant metric

Optimization Method

Adadelta

Adam

Long term goals

Unspervised learning (by predict and compare to the next t+n frame) of depth or optical flow

Train an Discriminator as supervisor