Interpreting extended, untrimmed videos by dividing them into meaningful action segments is fundamental for automated analysis of human activities. While modern models can represent long-range temporal dependencies, they often lack explicit mechanisms that encourage consecutive frames to follow coherent label progressions. We formulate segmentation as an energy-optimization problem composed of two parts: (i) a data term derived from frame-wise class likelihoods and (ii) a temporal regularization term that encourages coherent and semantically meaningful transitions between labels. Data costs are derived from a diffusion-based generative model (DiffAct) to capture action probabilities, while smoothness costs enforce semantic coherence by modeling valid transitions between action labels. The resulting discrete objective is optimized via α-β swap graph-cut moves. Results obtained on the GTEA dataset reveal that GBSO surpasses competitive baselines by providing more accurate segmentations and notably steadier temporal behavior, reflected in sharper action boundaries and reduced fragmentation. These outcomes demonstrate that embedding semantic transition structure within a data-centric segmentation approach leads to more reliable temporal predictions.

Keywords