how to find actioness score ?
Based on the spatio-temporal feature F
, we can conduct action localization in the spatial domain in a manner similar to
object detection in images. Inspired by the success of Cascade
RCNN [33], the proposed RAN contains two actionness stages
and one classification stage. Different from Cascade RCNN [33],
we have no bounding box annotations; hence, the proposed RAN
has no regression branch. In our method, for each stage, we apply Region Of Interest (ROI) pooling [18] method to extract the
region features for each proposal, and conduct binary classification on them. Based on the predicted actionness score, we select
the top N i proposal, i ∈ 1, 2, 3, for the next stage. For example, we apply the score C1 ∈ RN1×2 to select N2 boxes from
B1 to form B2, and the same process is applied in the second
stage