Temporal Action Localization (TAL) requires jointly optimizing action classification and temporal boundary regression. However, these two tasks favor conflicting feature representations: classification relies on semantically rich context, whereas localization demands fine-grained temporal details. Most existing one-stage TAL detectors decouple these tasks only at the prediction heads while sharing intermediate features, causing task interference and limiting boundary precision. To address this, we propose Text-Guided Decoupling for Temporal Action Localization (TGD-TAL), a text-conditioned framework that disentangles task-specific representations using video-derived textual priors. Specifically, we employ a Multimodal Large Language Model (MLLM) to generate descriptive text, which is encoded as semantic priors to guide feature modulation via a Text-Conditioned Feature Modulation (TCFM) module. Building upon a temporal feature pyramid, we design two complementary branches: the Text-Guided Classification Decoupling (TGCD) enhances category discrimination by attending to high-level contextual features, while the Boundary-Aware Text-Guided Regressor (BTR) refines temporal boundaries by leveraging low-level details and start- and end-aware textual cues via text–video cross-attention and gated fusion. Experiments show our method outperforms baselines, validating text priors’ effectiveness in reducing interference and sharpening boundaries.
原文链接
分享
引用
