Semi-iNat (Semi-Supervised iNaturalist)

此挑战集中在半监督的细粒度分类上，其中我们提供目标类别的标记数据以及来自目标和非目标类别的未标记数据。数据来自iNaturalist，这是一个社区驱动的项目，旨在收集生物多样性的观察结果。请参阅下面有关此挑战如何与FGVC和其他现有数据集上的先前的iNaturalist挑战相关的信息。
我们的数据集附带标准训练、验证和测试集。训练集包括:
来自810物种的标记图像，其中大约10% 的图像被标记。
未标记图像包含来自与标记图像 (类内) 相同的类集的未标记图像，以及来自与标记集 (类外) 不同的类集的图像。保证该物种在标签集中具有相同门级别的物种。这反映了一种常见的情况，在这种情况下，可以轻松获得图像的较粗分类标签。我们还为每个未标记的图像提供王国和门注释。

This challenge focuses on semi-supervised fine-grained classification, where we provide labeled data for target categories alongside unlabeled data from both target and non-target classes. The data is sourced from iNaturalist, a community-driven project dedicated to collecting biodiversity observation records. Please refer to the information below regarding how this challenge relates to FGVC and prior iNaturalist challenges on other existing datasets.

Our dataset includes standard training, validation, and test splits. The training set consists of:
Labeled images from 810 species, with approximately 10% of the total images being labeled.
Unlabeled images include both unlabeled samples from the same class set as the labeled images (in-class) and samples from classes distinct from the labeled set (out-of-class). All unlabeled in-class species are guaranteed to belong to the same phylum as those in the labeled set. This reflects a common real-world scenario where coarse-grained classification labels for images can be easily acquired. We also provide kingdom and phylum annotations for each unlabeled image.

搜集汇总

背景与挑战

背景概述

Semi-iNat是一个用于半监督细粒度分类的数据集，数据来自iNaturalist项目，包含810个物种的标记和未标记图像，其中标记图像约占10%，未标记数据涵盖类内和类外样本，并提供王国和门级别的注释。该数据集由马萨诸塞大学阿默斯特分校于2021年发布，旨在支持生物多样性观察中的分类挑战。