Table 2.
Method | Classification |
Detection |
||||||
---|---|---|---|---|---|---|---|---|
P | R | F-score | △ | P | R | F-score | ||
One-stage | SCNN1 | 0.691 | 0.651 | 0.670 | 7.6% | 0.747 | 0.768 | 0.757 |
UTurku (Björne et al., 2013A) | 0.732 | 0.499 | 0.594 | 0.858 | 0.585 | 0.696 | ||
NIL_UCM (Bokharaeian and Dıaz, 2013) | 0.535 | 0.501 | 0.517 | 0.608 | 0.569 | 0.588 | ||
Two-stage | SCNN2 | 0.725 | 0.651 | 0.686 | 1.6% | 0.775 | 0.769 | 0.772 |
Kim et al., (2015) | – | – | 0.670 | – | – | 0.775 | ||
FBK-irst (Chowdhury and Lavelli, 2013) | 0.646 | 0.656 | 0.651 | 0.794 | 0.806 | 0.800 | ||
WBI (Thomas et al., 2013) | 0.642 | 0.579 | 0.609 | 0.801 | 0.722 | 0.759 |
Notes. SCNN1 denotes our SCNN-based one-stage method and SCNN2 denotes our SCNN-based two-stage method. Δ denotes the performance improvement of SCNN1 over UTurku, and SCNN2 over that of Kim et al. (2015). The boldfaced numerals are the highest values in the corresponding column.