Zanim

Zanim ('zero-shot animals') is a dataset specifically designed to benchmark few-shot and multi-modal image classification; each animal in the dataset has a succinct description of its appearance, and, as a result, it is desirable to create models that can leverage this information to improve few (or zero)-shot image-classification. 'data.gz' - contains all the images of the dataset including the file structure specified in the json 'am3.pth.tar' - a checkpointed version of the AM3 model used in our research 'fumi.pth.tar' - a checkpoint of the FuMI model used in our research 'cropped-images-integers.hdf5' - cropped and downscaled images (128x128) ideal for viewing in a notebook 'image-embedding-resnet-152.hdf5' - embeddings of each image in the dataset using a resnet-152 model 'zanim.json' - a json file containing the descriptions and additional information required for modellingZanim（即“零样本动物（zero-shot animals）”）是专为少样本与多模态图像分类基准测试设计的数据集。该数据集内的每一类动物均配有简洁精准的外观描述，因此开发能够利用此类信息优化少样本（或零样本）图像分类性能的模型，成为了极具研究价值的方向。以下为数据集配套文件说明： - data.gz：包含数据集全部图像，且附带JSON文件中指定的文件目录结构 - am3.pth.tar：本研究所用AM3模型的检查点文件 - fumi.pth.tar：本研究所用FuMI模型的检查点文件 - cropped-images-integers.hdf5：经过裁剪与下采样至128×128分辨率的图像，适配在交互式笔记本环境中查看 - image-embedding-resnet-152.hdf5：使用ResNet-152模型对数据集内各图像提取得到的特征嵌入文件 - zanim.json：包含建模所需的图像描述与额外相关信息的JSON文件