imageomics/fine-grained-challenges
收藏资源简介:
Fine-Grained Challenges数据集收集了视觉相似的动物图像组,用于测试生物图像分类器。它在一个Lance数据集中结合了图像、分类学、来源和来自三个BioCLIP家族模型的冻结嵌入。数据集包含三个挑战组:Peromyscus(鹿鼠及其近缘种,共1,737行,其中1,332行有物种标签,405行只有属或更高分类标签,涵盖37个物种)、Ixodidae(硬蜱,共4,499行,其中4,423行有物种标签,76行只有属或更高分类标签,涵盖173个物种)和zebra(马属中的斑马物种,共397行,全部有物种标签,涵盖4个物种)。总计6,633行数据,其中6,152行有物种标签,481行只有属或更高分类标签,涵盖214个物种。数据集还包括305个仅解析到属的相机陷阱图像,这些图像未包含在BioCLIP模型的训练数据中,可用作开放式、类似野外的检查集。数据集支持细粒度图像分类、零样本分类、基于冻结嵌入的小样本分类等任务。数据来源包括TreeOfLife-200M和LILA BC等集合,图像以WebP格式存储,最大边长为720像素,并附有预计算的嵌入向量。数据集发布在CC0-1.0许可下,但每个图像保留其原始许可。
Fine-Grained Challenges collects focused groups of visually similar animals for testing biological image classifiers. It combines images, taxonomy, provenance, and frozen embeddings from three BioCLIP-family models in one Lance dataset. The dataset includes three challenge groups: Peromyscus (deermice and close relatives, with 1,737 rows, of which 1,332 are species-labeled, 405 are genus or higher, covering 37 species), Ixodidae (hard ticks, with 4,499 rows, of which 4,423 are species-labeled, 76 are genus or higher, covering 173 species), and zebra (zebra species in genus Equus, with 397 rows, all species-labeled, covering 4 species). In total, there are 6,633 rows, with 6,152 species-labeled rows, 481 genus or higher rows, covering 214 species. The dataset also includes 305 genus-only camera-trap images that were not included in the training data of the BioCLIP models and can be used as an open-ended, field-like inspection set. It supports tasks such as fine-grained image classification, zero-shot classification, few-shot classification using frozen embeddings, and more. Data sources include TreeOfLife-200M and LILA BC collections, with images stored in WebP format (maximum edge of 720 pixels) and precomputed embeddings. The dataset is released under CC0-1.0, but each image retains its original license.




