BAM-DETR: Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos

Temporal sentence grounding aims to localize moments relevant to a languagedescription. Recently, DETR-like approaches achieved notable progress bypredicting the center and length of a target moment. However, they suffer fromthe issue of center misalignment raised by the inherent ambiguity of momentcenters, leading to inaccurate predictions. To remedy this problem, we proposea novel boundary-oriented moment formulation. In our paradigm, the model nolonger needs to find the precise center but instead suffices to predict anyanchor point within the interval, from which the boundaries are directlyestimated. Based on this idea, we design a boundary-aligned moment detectiontransformer, equipped with a dual-pathway decoding process. Specifically, itrefines the anchor and boundaries within parallel pathways using global andboundary-focused attention, respectively. This separate design allows the modelto focus on desirable regions, enabling precise refinement of momentpredictions. Further, we propose a quality-based ranking method, ensuring thatproposals with high localization qualities are prioritized over incompleteones. Experiments on three benchmarks validate the effectiveness of theproposed methods. The code is available athttps://github.com/Pilhyeon/BAM-DETR.