Query Strategies for 3D model

[jonaskordt]

Query strategies are currently not working for 3D models.

Random:

• Samples a random scan without a slice, which results in labeling slice 0 of that scan
• Could be easily fixed by checking which slices of that scan are unlabeled and sampling a random one of those

Uncertainty:

• Breaks when using a 3D model because the predictions have an additional dimension
• Could be fixed by splitting the predictions along the slice dimension and adjusting the corresponding case_id to contain the correct slice number

Interpolation:

• Same as uncertainty because it (currently) uses uncertainty sampling to select the interpolation blocks

Representativeness:

• The feature vectors used to rank representativeness have an additional spatial dimension
• They could be split along this dimension and each slice could be assigned the closest feature-vector-slice
• This would result in multiple neighboring slices getting the same value, which would be fine, since neighboring slices are mostly pretty similar