vCLIMB

vCLIMB是一个专注于视频动作分类任务的持续学习基准，由沙特阿拉伯国王阿卜杜拉科技大学创建。该数据集整合了UCF101、Kinetics和ActivityNet三个大型视频数据集，通过定义固定的任务分割，支持10和20个任务序列的持续学习评估。vCLIMB特别关注未修剪视频的处理，模拟真实世界中从连续视频流中学习的场景。数据集的创建过程中，重新定义了记忆大小以适应视频数据的特性，并引入了时间一致性正则化来优化模型性能。vCLIMB的应用领域包括视频动作识别，旨在解决模型在持续学习新任务时如何保持对旧任务性能的问题。

vCLIMB is a continual learning benchmark specialized for video action classification tasks, developed by King Abdullah University of Science and Technology (KAUST) in Saudi Arabia. This dataset integrates three large-scale video datasets: UCF101, Kinetics, and ActivityNet. By defining fixed task splits, it supports continual learning evaluations with 10 and 20 task sequences. vCLIMB places particular emphasis on untrimmed videos, simulating the real-world scenario of learning from continuous video streams. During the dataset construction process, the memory size was redefined to accommodate the unique characteristics of video data, and temporal consistency regularization was introduced to optimize model performance. Targeted at video action recognition applications, vCLIMB aims to address the challenge of how models maintain performance on previously learned tasks while continually learning new ones.