yamins81 / eccv12 Goto Github PK

@jaberg, do you think that the boosting in rounds should be implemented as a:

1. function that creates multiple experiments in series, with e.g. J different but related exp_keys (where J is the number of boost rounds)?

2. a BoostedExperiment class that maybe keeps one exp key but basically calls the standard experiment J times?

3. A bandit algo wrapper that wraps any given bandit algo and creates a new one which basically implements the relevant stuff?

4. Something totally different?

The FG11 paper used input rescaling to boost performance. Can we make the input resolution a parameter of the search space? What downsampling possibilities should we choose?