plantseg-lookalike
收藏资源简介:
PlantSeg Subtle Disease Look-Alike Pairs(植物分割细粒度疾病相似对)数据集是一个包含45,393对视觉上难以区分的植物疾病图像对的基准数据集,专门针对同一作物、同一植物器官上两种不同但外观极其相似的疾病症状。该数据集从PlantSeg v3中挖掘而来,每个图像对都经过严格的网络验证,包含来自农业文献的引述,确认这些疾病在实际诊断中容易混淆。构建过程采用四层互补筛选标准:通过DINOv3模型在真实病变掩码上计算图像对余弦相似度(要求≥0.80),确保同一植物器官和取景方式一致,并检索网络验证引文。数据集经过重复图像完整性检查,涵盖45种不同的疾病对和19种作物,其中大豆、小麦、玉米为主要作物。每个数据样本包含15个字段,如图像对缩略图、疾病名称、作物类型、相似度分数等。该数据集适用于细粒度疾病混淆性研究、难负例挖掘、零样本和细粒度分类器的失败模式分析等任务,基于CC BY 4.0许可,使用时需引用本数据集和原始PlantSeg v3数据集。
The PlantSeg Subtle Disease Look-Alike Pairs dataset is a benchmark dataset containing 45,393 pairs of visually indistinguishable plant disease image pairs, specifically targeting two different but highly similar disease symptoms on the same crop and plant organ. It is mined from PlantSeg v3, with each image pair rigorously web-verified and including citable quotes from agricultural literature confirming that the diseases are easily confused in actual diagnosis. The dataset construction employs a four-layer complementary screening criteria: computing cosine similarity (≥0.80) on real lesion masks using the DINOv3 model, ensuring the same plant organ and consistent framing, and retrieving web verification citations. It undergoes duplicate image integrity checks, covering 45 distinct disease pairs and 19 crops, with soybean, wheat, and corn as the main crops. Each data sample includes 15 fields, such as image pair thumbnails, disease names, crop type, similarity score, etc. The dataset is suitable for tasks like fine-grained disease confusion research, hard negative mining, failure mode analysis of zero-shot and fine-grained classifiers, and is based on CC BY 4.0 license, requiring citation of both this dataset and the original PlantSeg v3 dataset when used.




