Detecting temporal actions in long and untrimmed videos is a challenging and important field in computer vision. Generating high-quality proposals is a key step in temporal action detection. A high-quality proposal usually contains two main characteristics. One is the temporal overlaps between proposals and action instances should be as large as possible. The another one is the number of generated proposals should be as few as possible. Inspired by the similarity comparison in face recognition and the similarity of action in same action segment, we design a module to compare the similarity for visual features extracted from visual feature encoder. We find out time points where the similarity of features changes shapely to generate candidate proposals. Then, we train a classifier to evaluate the candidate proposals whether contains or not contains action instances. The experiments suggest that our method outperforms other temporal action proposal generation methods in THUMOS-14 dataset and ActivityNet-v1.3 dataset. In addition, our method still outperforms other methods when using different visual features extracted from different networks.

在较长且未修剪的视频中检测时间动作是计算机视觉中一个具有挑战性且重要的领域。生成高质量的建议是临时动作检测中的关键步骤。高质量的建议通常包含两个主要特征。一是提案和行动实例之间的时间重叠应尽可能大。另一个是所生成提案的数量应尽可能少。受人脸识别相似性比较和同一动作段中动作相似性启发，我们设计了一个模块来比较从视觉特征编码器提取的视觉特征的相似性。我们找出特征相似度发生形变以生成候选建议的时间点。然后，我们训练分类器来评估候选提案是否包含动作实例。实验表明，在THUMOS-14数据集和ActivityNet-v1.3数据集中，我们的方法优于其他时间行动建议生成方法。此外，当使用从不同网络提取的不同视觉特征时，我们的方法仍然优于其他方法。