Official page for Matlab system parameters, CVPR 2014 (c. October 2014).
Any changes to the matlab parameters need to be updated here, with a corresponding update to the changelog.
params
params =
smoothing_variance_2d: 0.0100
noise_variance_2d: 0.3480
position_mean_2d: [265 198.5000]
position_variance_2d: 15609
rate_variance_2d: 2.2500
smoothing_variance: 0.0035
noise_variance: 0.3480
position_mean: [3x1 double]
position_variance: 1.2987e+03
rate_variance: 0.2652
index_delta: 2
index_delta_2d: 2
perturb_smoothing_variance: 2.7745e-06
perturb_rate_variance: 4.3973e-04
perturb_position_variance: 0.6356
perturb_scale: 2.4900
markov_blanket_radius: 2000
model_type: 3
num_views: 9
ml_block_size: 1000
ml_markov_order: 1000
ll2_spacing: 4
tube_radius: 1
posterior_markov_order: 10
posterior_block_size: 100
Most recent items first.
ll2_spacing: 2 -> 4. Changed during an attempt to speed up posterior sampling. Didn't have much affect, positive or negative, because the bottleneck was the size of the data, not the output indices. Still some room for experimenting here...
posterior_markov_order (new): 10. Set heuristically. Incurs some error, unclear if significant, but I'm feeling optimistic at the mmoment. Slightly better than zero; need around 500 for error to decrease to zero (possibly no markov benefit at this size).
posterior_block_size (new): 100. Did grid search from 10 to 1000. Roughly optimal speed-wise; better than 80 or 120.
Initial import.
Most prior parameters came from training on August 16, 2013, with the exception of position_variance
, which was modified when we added the position_mean parameter. The old position_variance
was derived from an assumed mean of (0,0,0); after learning a mean, the variance naturally dropped. The perturb parameters are actually slightly different from the August 16 results; not sure how the deviation aros, but they're still in the ballpark. Retrain soon.
index_delta
and index_delta_2d
were hand-set heuristically.
markov_blanket_radius
is basically "infinity"; all parent points are used when computing branch point. This is a good candidate for tuning; IIRC, I set it this high during testing and never changed it back. Lowering should improve attach()
runtimes.
model_type
is OU-perturb-model (i.e. 3).
ml_block_size
and ml_markov_order
were hand-set heuristically to be a good balance between speed and approximation accuracy (IIRC). Possibly some room for improvement here.
ll2_spacing
is a new parameter, the sampling period when evaluating the "second" likelihood, i.e. the pixel likelihood. Arbitrary, untested.
tube_radius
also pertains to the pixel likelihood. Its the radius of the tubes rendered by opengl. Hand-picked, relatively untested (but qualitatively reasonable.).