Closed qxcv closed 8 years ago
Incomplete list of where those variables appear:
train/mark_scales.m
1:function [dataset, tsize] = mark_scales(dataset, subposes, step, template_scale, other_scales)
11:% other datasets so that there is more data available to calculate .tsize
37:tsize = round(smallish_scale / step);
mark_scales
is the only relevant place where there is still a tsize
. I don't think I'll remove it, since I may find that tsize
becomes useful on other datasets. For now, however, it's not used at all.
They only made sense when we didn't have explicit sizes for our poselets (derived during training). Now they are just weird cruft which is probably hurting performance due to the imprecision they introduce. I should remove them entirely, replacing each with
cnn.window
orcnn.window / cnn.step
depending on context.