Open pv33 opened 10 months ago
pv33
commented
10 months ago Note: Only the location is pulled out from the PuzzleTemplate information into a separate member variable in Template. Seems like it might be OK as long as Template does the right thing. The right thing may be established by a flag per above discussion.
pv33
commented
10 months ago A board has a flag to determine whether it will apply updates to its pieces or simply associate with the measurement.
Pieces have their own update function that establishes how to update their content. The member function will recurse to feature descriptor updating.
Not entirely sure that it is correct, but the current approach generates the associations, pushes new feature vector to estimate, and then pushes labels of estimate to the measurement. Plotting the measurement gives the current state. Plotting the estimate will give the original/source state. How proper is this? May ultimately need estimate to be updated. Rolling with code as is until I get a better grasp on the API.
For different properties, will need to overload the puzzle piece class (create a hierarchy of implementations). Another option is to add in an update descriptor / configuration parameter for simple cases (not done, but could be). This kind of parameter would enrich the board level parameter. Keep this in mind too.
Options
The feature descriptor consists of information pulled from the piece state. When performing data association, what needs to change/be updated? We have the following options regarding the original template piece:
1) Puzzle piece raw data remains fixed. 2) Puzzle piece raw data is overwritten. 3) Puzzle piece raw data is blended / merged / averaged / etc.
We also have the following options regarding the feature descriptor:
1) Keep original feature vector. 2) Replace original feature vector. 3) Blend feature vectors somehow (up to feature vector format I suppose). 4) Recalculate feature vector (usually after blending raw data).
Ideal for Puzzle
Ideally we would have the original template, the latest matched template, and continued frame to frame matching with the latest matched that is ultimately associated to the original template.
That has a solution to latest association, plus a latest to current association. The association to the solution is for knowing where to place the pieces. The association of latest to current is to keep track of where things are more generally and how they move around. Moved is indirect work or direct work? Move achieves progress or not in terms of puzzle solving?
Solution Board
Solution piece correspondence would update feature vector but not original source data. It would definitely update the labels. It would be aware of "label constancy" (new things given high number past current labels). That means also checking for non-associated and having some kind of new piece label function.