发表机构
Alibaba Group; Beihang University; Fudan University（阿里巴巴集团; 北京航空航天大学; 复旦大学）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
本文提出一种可控生成式视频时序定位方法，通过轻量级置信度头在解码过程中为候选区间打分，支持排序、阈值筛选和拒绝，无需外部验证器，在OMTG-Bench上显著提升召回率。
AI 中文摘要
视频时序定位通过定位自然语言描述的事件，支持视频搜索、内容审查和自动编辑等应用。然而，现有的生成模型通常输出时间戳，而没有明确的区间级置信度分数来指导候选选择。我们将候选生成与接受分离，在原始解码过程中对单个区间进行评分。一个轻量级置信度头读取池化后的解码器状态，提供经过训练的显式分数用于区间选择。离线验证器分数在固定候选序列上监督该置信度头，时间重叠标签在强化学习期间使其适应当前的滚动输出。GT锚定的候选池监督和集合级优化训练生成器。所得分数支持排序、基于阈值的筛选和拒绝，而无需在推理时调用外部验证器。在固定的OMTG-Bench候选池上，置信度在10%的全局返回预算下，将查询宏平均Recall@0.5从9.95%提升至14.42%，在25%预算下从26.48%提升至31.12%。连续分数使下游应用能够调整返回预算或接受阈值，以匹配其精确率-召回率偏好，而无需重新生成候选区间。
英文摘要
Video temporal grounding supports applications such as video search, content review, and automated editing by localizing events described in natural language. Yet existing generative models typically output timestamps without explicit interval-level confidence scores to guide candidate selection. We separate candidate generation from acceptance by scoring individual intervals within the original decoding pass. A lightweight confidence head reads pooled decoder states, providing an explicit score trained for interval selection. Offline verifier scores supervise the head on fixed candidate sequences, and temporal-overlap labels adapt it to current rollouts during reinforcement learning. GT-anchored candidate-pool supervision and set-level optimization train the generator. The resulting scores support ranking, threshold-based selection, and rejection without invoking an external verifier at inference. On a fixed OMTG-Bench candidate pool, confidence raises query-macro Recall@0.5 from 9.95% to 14.42% over generation order at a 10% global return budget, and from 26.48% to 31.12% at a 25% budget. The continuous scores let downstream applications adjust return budgets or acceptance thresholds to match their precision-recall preferences, without regenerating candidate intervals.
Comments22 pages, 7 figures; includes appendix
