Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
M.Sc. Thesis: Controlling tree foliage in images using deep networks
Controlling the amount and appearance of foliage is important for hallucinating seasonal changes.
Controlling the amount and appearance of foliage is important for hallucinating seasonal changes. Appearance changes can be modelled as local affine transformations of colour space [1,2]. However, colour transfer alone is not sufficient for controlling the amount of foliage [2,3]. Recent methods have used a combination of colour and texture transfer from foliage of an exemplar tree to a barren tree [3]. They require registered pair of barren and green exemplars and dense correspondences between exemplar tree and barren tree. These requirements make the methods almost impractical for real world usage.
We want to use deep learning to add/remove foliage from trees. We will train a convolutional neural network (CNN) to estimate the probability of leaf at each pixel of an image of an empty tree. Ideally this probability map should be high near small branches and low near the trunk of the tree. Actual leaf addition can be accomplished by texture synthesis. We will train the CNN on continuous levels of foliage {10, 20 .. 90%} so that the network can synthesise partially and fully covered trees.
Controlling the amount and appearance of foliage is important for hallucinating seasonal changes. Appearance changes can be modelled as local affine transformations of colour space [1,2]. However, colour transfer alone is not sufficient for controlling the amount of foliage [2,3]. Recent methods have used a combination of colour and texture transfer from foliage of an exemplar tree to a barren tree [3]. They require registered pair of barren and green exemplars and dense correspondences between exemplar tree and barren tree. These requirements make the methods almost impractical for real world usage. We want to use deep learning to add/remove foliage from trees. We will train a convolutional neural network (CNN) to estimate the probability of leaf at each pixel of an image of an empty tree. Ideally this probability map should be high near small branches and low near the trunk of the tree. Actual leaf addition can be accomplished by texture synthesis. We will train the CNN on continuous levels of foliage {10, 20 .. 90%} so that the network can synthesise partially and fully covered trees.
Not specified
Gaurav Chaurasia (gaurav.chaurasia@disneyresearch.com)
Torsten Sattler (sattlert@inf.ethz.ch)
Martin Oswald (martin.oswald@inf.ethz.ch)
Gaurav Chaurasia (gaurav.chaurasia@disneyresearch.com) Torsten Sattler (sattlert@inf.ethz.ch) Martin Oswald (martin.oswald@inf.ethz.ch)