Advanced 2D/3D segmentation

Question 1

When does voxelization work?

Answer 1

• small images

- dense images

Question 2

What is octnet and what are its benefits?

Answer 2

A 3D segmentation scheme that does 3D convolutions irregularly on an image based on the important locations. Efficient (memory, computation) on large sparse 3D images and resolution invariant (not really, but does not improve much on higher resolution).

Question 3

What is multi-view?

Answer 3

2D-sampling of 3D images. Find most important view points of a 3D object and do 2D segmentation, then recombine into a 3D segmented object. Problem is finding views that maximise area coverage.

Question 4

Mention some situations where multi-view is a poor solution

Answer 4

• Large complex point cloud
• Dense point-cloud
• Noisy/sparse point cloud

Question 5

Give a brief explanation of PointNet

Answer 5

A deep learning 3D model for classification and segmentation. Works directly on point cloud in contrast to most other algorithms. Inputs n 3D points and outputs score based on 1024 points (max-pooled) (not sure how output looks).

Question 6

Mention four different ways of doing 3D segmentation

Answer 6

Using voxels/Octrees (VoxNet/OctNet), Multi-view (multiple 2D images), directly from point cloud (PointNet), Kd-trees (Kd-networks), graphs (graph convolutions).