20
I. Athanasiadis et al.
Fig. 1. The architecture combining cascade with EA-CNNC .
produced by the MSER algorithm occupying an area larger than 15
2 pixels or
have their aspect ratio falling outside the [0.5, 2] interval, are filtered out.
Results: To evaluate the impact of each method on the detection performance,
we report the standard COCO [6] metrics in Table 1. Notably, utilising the cascade approach seems to boost the metrics of the strict IoU threshold. Additionally, improved performance is achieved, in terms of AP S , when edge anchors are
used combined with high-quality cascade classifiers. Moreover, the results shown
in Fig. 2, referring to screws, are indicative of the edge anchors’ capability to
detect significantly small objects.
Table 1. Object detection results on PC-Tower dataset in all classes (all its components).
Method
Cascade AP0.5:0.95 AP0.5 AP0.75 APS AR0.5:0.95 ARS
Baseline
–
40.3
69.8 39.5
23.1 47.3
28.0
39.5
66.2 40.1
23.6 48.8
29.0
EA-CNNP
–
38.7
68.2 38.0
22.9 46.9
28.8
40.9
70.5 42.3
24.6 48.2
30.1
EA-CNNC
–
39.6
69.5 40.1
25.5 47.2
31.4
40.6
69.5 41.9
25.2 48.4
32.3