where
and represent the weighing factor and focusing parameter, respectively.
The other notations refer to Equation . In practice, and are separately
set to 0.25 and 2 , according to the parameter setting in [48]. The
final loss function for ROI Head is obtained by adding the above loss
and :
The losses of ROI Head are accumulated and back-propagated to train the
proposed framework.
3.4 Pi-Index: the strategy of estimating class
prior
In the Faster RCNN framework, the PU loss is usually applied to the
anchor-based RPN [34], [49]. In other words, classification and
regression are performed separately for each anchor in RPN. Incomplete
annotations should thus be converted from annotated boxes to anchors.
There are lots of positive anchors assigned negative labels as
background because a part of the targets lacks their annotations. For
the PU classification loss of RPN, the class prior is defined as the
percentage of both correctly and incorrectly labeled positive anchors.
A. Background: the existing
methods
Zhao et al. considered the class prior as a hyper-parameter and
determined it by grid search based on a validation set [49]. The
estimation granularity of the class prior relies on the interval of the
grid search.
Table 1: Performance (AP) of the method in [34] with
different confidence thresholds (0.1-0.5) and annotation percents (0.3,
0.5, 0.7 and 1).