Flowers-620
收藏资源简介:
Flowers-620数据集是由Cornell University和NEC Labs America合作创建的细粒度花卉分类数据集,包含620种花卉的20,211张图像。该数据集通过深度度量学习和人类参与的迭代框架进行数据集引导,旨在解决细粒度视觉分类中的训练数据缺乏、类别众多和类内高变异与类间低变异的问题。数据集的创建过程中,利用模型生成的置信度高的图像进行人工标注,筛选出真实正例和困难负例,不断扩充和优化数据集。Flowers-620数据集主要应用于细粒度花卉分类研究,通过该数据集的引导,能够提升模型的分类性能,达到当前最先进的技术水平。
The Flowers-620 dataset is a fine-grained flower classification dataset co-developed by Cornell University and NEC Labs America, comprising 20,211 images spanning 620 flower species. This dataset is built upon a dataset-guided framework that integrates deep metric learning and human-in-the-loop iteration, with the goal of addressing core challenges in fine-grained visual classification: insufficient training data, large-scale category sets, high intra-class variability, and low inter-class variability. During its construction process, manually annotated high-confidence images generated by models were used to screen true positive samples and hard negative samples, enabling continuous expansion and optimization of the dataset. The Flowers-620 dataset is primarily utilized for fine-grained flower classification research. Leveraging this dataset for model guidance can effectively improve classification performance, achieving the current state-of-the-art level.

- 1Fine-grained Categorization and Dataset Bootstrapping using Deep Metric Learning with Humans in the Loop美国国家航空航天局 · 2016年



