[Work Log] Refactor: one-model-per-view likelihood
Implementing one-object-per-view in likelihood_server
Overloaded bd_mv_likelihood' evaluate and dump functions to receive a sequence of renderables.
Write a "wrap_all_As_silhouette" to, well, wrap all renderables in silhouette renderers.
re-wrote message-to-curve function to decode the message to a vector-vector-vector, and convert that to a vector-of-gl_curves.
Compiling work from last 24 hours.
Done.
- Generate a multi-model message from matlab (dummy)
- test message in likelihood_server_2.cpp
- generate random samples from wacv data.
Doing dummy message test.
At first, accidentally sent to old implementation... and it didn't barf! This is unsettling, because the message format has changed significantly. Did I ever recompile the mex files?
Nope. got some mex compile errors to deal with.
Friday
okay, sending a random model using the new one-per-view message system isn't crashing.
Questions
- the likelihood looks high, considering it's a ranomd model . is a null model better? is the ground-truth model better?
- visualize the received message; is each view different?
Trying the null model from 1., the mex file crashed matlab. Empty curvesset isn't handled.
trying all-zeros. Server crashed -- coincident points aren't handled gracefully. Fixed (now it returns -inf for log likelihood).
issues server: error creating gl_curves causes client disconnect. fix exception handling client: doesn't handle server crashes gracefully. adding timeout client: doesn't handle empty curveset
Next:
- random sample from multi-view posterior (instead of single max posterior)
- get likelihood from multiple samples -- what is the variance of the MC estimator?