LPR: learning point-level temporal action localization through re-training

Abstract Point-level temporal action localization (PTAL) aims to locate instances in untrimmed videos with only one timestamp annotation for each instance. Existing methods adopt the localization-by-classification paradigm boundaries class activation map (TCAM) by thresholding, also known as TCAM-based method. However, are limited gap between classification and tasks, since TCAM is generated a network. To address this issue, we propose re-training framework PTAL task, LPR. This consists of two stages: pseudo-label generation re-training. In stage, feature embedding module based on transformer encoder capture global context features optimize pseudo-labels’ quality leveraging point-level annotations. LPR uses above pseudo-labels supervision network rather than generating TCAMs. Furthermore, alleviate effects label noise pseudo-labels, joint learning (JLCM) stage. contains sub-modules that simultaneously predict categories guided jointly determined clean set training. The proposed achieves state-of-the-art performance both THUMOS’14 BEOID datasets.