Supplemental Material for Contextual Proposal Network for Action Localization

This part discusses the benefits of our boundary scoring mechanism. Table 1 summarizes the ablation study to assess our model’s various configurations. Please notice that we assume that each model configuration in Table 1 equips with our complete feature enhancing mechanism. 1 Table 1 has six model configurations, and only the first configuration, i.e., baseline-BS, is implemented without using RNNs and directly predicts the probabilities of p and p akin to BSN yet with the enhanced feature F. In contrast, the other rows are RNN-based boundary predictions using the hidden states of RNNs.