FACETS OOD Detection
收藏资源简介:
FACETS OOD Detection数据集是由都灵理工大学的FRESCO研究线开发的,旨在通过深度学习工具分析社交媒体上的大量个人资料图片。该数据集模拟了一个预训练的图像标记模型,需要在开放世界假设下操作。数据集主要用于场景分类任务,以Places365数据集作为ID集,从ImageNet中抽取OOD样本。数据集的创建涉及自动和手动标记技术,以及基于WordNet的语义相似性度量,以确定哪些类应被视为ID、近OOD或远OOD。该数据集的应用领域包括社交媒体图像分析,旨在解决模型在面对未知样本时的过拟合和错误预测问题。
The FACETS OOD Detection Dataset was developed by the FRESCO research line at Politecnico di Torino. It aims to analyze large volumes of profile pictures on social media using deep learning tools. This dataset simulates a pre-trained image tagging model that must operate under the open-world assumption. The dataset is primarily used for scene classification tasks, taking the Places365 dataset as the in-distribution (ID) set and extracting out-of-distribution (OOD) samples from ImageNet. Its creation involves both automatic and manual annotation techniques, alongside WordNet-based semantic similarity metrics to classify which classes qualify as ID, near-OOD, or far-OOD. The application scope of this dataset covers social media image analysis, and it is designed to address the issues of model overfitting and erroneous predictions when encountering unknown samples.

- 1Toward a Realistic Benchmark for Out-of-Distribution Detection都灵理工大学 · 2024年



