FSL

具有少镜头学习和开放集识别问题的新音频数据集。数据集由34个类组成。其中24个被认为是已知类 (要分类的类)，其余10个未知类 (要拒绝的类)。关于类及其分布的更多信息可以在论文中找到。由于数据集的配置，可以创建两种不同的实验方案。将对24个已知类进行分类。根据训练中存在的未知类的数量，开放性的值可以是0、0.04或0.09。将24个已知类分为8组3。根据训练中存在的未知类的数量，开放性值可以是0、0.13或0.39。数据集基于转移学习技术提供了两个可能的基线。用作特征提取器的网络是L3Net和YAMNet。

This is a novel audio dataset designed for few-shot learning and open-set recognition problems. The dataset contains 34 classes in total, with 24 classified as known classes (classes to be classified) and the remaining 10 as unknown classes (classes to be rejected). More details regarding the classes and their distributions can be found in the associated paper. Two distinct experimental protocols can be constructed based on the dataset's configuration. For the first protocol, classification tasks are conducted on the 24 known classes, where the openness values can be 0, 0.04, or 0.09 depending on the number of unknown classes included during training. For the second protocol, the 24 known classes are divided into 8 groups each containing 3 classes, and the openness values can be 0, 0.13, or 0.39 depending on the number of unknown classes used in training. The dataset offers two viable baselines built upon transfer learning techniques, with L3Net and YAMNet serving as the feature extractors.

搜集汇总

背景与挑战

背景概述

FSL是一个专为少镜头学习和开放集识别设计的音频数据集，包含24个已知类别和10个未知类别。它支持多种开放性配置的实验方案，并提供了基于L3Net和YAMNet的迁移学习基线。