4.4.2 Detection results under 0.3 annotation percentage
In this section, the APC is set to
0.3, which means that 70% of the labels in the IDID dataset are
randomly removed. The aim is to observe the detection performance of
each method when labels are severely missing and to verify the
effectiveness of the algorithm proposed in this paper under such
conditions.
In Tables 2-5, each detector’s
performance deteriorated as the APC decreased. Contrary to the results
of the previous experiment, Table 5 shows that the AP metric of YOLO v5
is lower than that of D-YOLO v4. It indicates that YOLO v5 is not
suitable for the situation where there is a severe lack of labels.
Nevertheless, the two-stage Faster RCNN surpasses other PN-based methods
with an AP of 58.57%, an AP@0.75 of 70.41%, and an AP@0.5 of 81.02%,
which is consistent with previous experiments. According
Table 5: Detection results of our method and the baseline
method with (APC=0.3) in the training process.