Pollen detection dataset

Overview This dataset is about corbicular pollen loads as foraged by honey bees (apis mellifera). Pollen were obtained by using pollen traps and show a wide range of colors. A total of 12568 (training) and 1629 (validation) pollen were annotated on 64 (training) and 5 images (validation), respectively. The images were partly from photographs and partly from scanned pollen. From these images, 96x96 pixel images were randomly cropped and augmented. Suitable masks were also created and saved together with the corresponding images and csv files. These cropped images form the given data set. File structure train/ train/imgs/ 495k images contain a total of 1410991 pollen. Images have shape (96px * 96px * 3 channels). Details on augmentation can be found in README.md. ~20% of the images show no pollen. Mathematically, each labeled pollen appears on 112 images in the data set. train/imgs/masks For each image exists a binary mask suitable for U-net training. The mask is not a real segmentation, but a white circle marking the center of the pollen. Each white circle has a black border that guarantees that no white circles overlap. This helps the U-net to learn the separation of the pollen segmentations, which simplifies the detection of blobs or local maxima on the output map of the U-net. train/imgs/coords For each image exists a .csv file with the annotated pollen centers. val/ val/imgs/ 50k images contain a total of 284125 pollen. Images have shape (96px * 96px * 3 channels). Details on augmentation can be found in README.md.. ~20% of the images show no pollen. Mathematically, each labeled pollen appears on 108 images in the data set. val/imgs/masks For each image exists a binary mask suitable for U-net training. The mask is not a real segmentation, but a white circle marking the center of the pollen. Each white circle has a black border that guarantees that no white circles overlap. This helps the U-net to learn the separation of the pollen segmentations, which simplifies the detection of blobs or local maxima on the output map of the U-net. val/imgs/coords For each image exists a .csv file with the annotated pollen centers.数据集概览 本数据集聚焦西方蜜蜂（Apis mellifera）花粉篮采集的花粉团。花粉样本通过花粉捕集器获取，样本色彩跨度广泛。数据集共包含标注花粉团12568份（训练集）与1629份（验证集），分别对应64张（训练集）与5张（验证集）原始图像。原始图像部分来自实拍照片，部分来自扫描的花粉样本。我们从上述图像中随机裁剪出96×96像素的图像并进行数据增强，同时生成配套二值掩码，与对应图像及CSV文件一同保存，最终裁剪后的图像构成该数据集。 文件结构 训练集目录（train/） train/imgs/：包含49.5万张图像，总计涵盖1410991个花粉团。图像分辨率为96像素×96像素×3通道。数据增强的详细说明可参阅README.md文件。约20%的图像未包含任何花粉团。从统计数学层面而言，数据集中每一个被标注的花粉团会出现在112张图像中。 train/imgs/masks/：为每张图像生成适用于U-net（U-Net）训练的二值掩码。该掩码并非真实的实例分割掩码，而是以白色圆圈标记花粉团的中心，每个白色圆圈均带有黑色边框，确保白色圆圈之间无重叠。这一设计可帮助U-net学习花粉团的分割边界，简化U-net输出特征图上的团块检测或局部极大值提取流程。 train/imgs/coords/：为每张图像生成包含标注花粉团中心坐标的CSV文件。 验证集目录（val/） val/imgs/：包含5万张图像，总计涵盖284125个花粉团。图像分辨率为96像素×96像素×3通道。数据增强的详细说明可参阅README.md文件。约20%的图像未包含任何花粉团。从统计数学层面而言，数据集中每一个被标注的花粉团会出现在108张图像中。 val/imgs/masks/：为每张图像生成适用于U-net（U-Net）训练的二值掩码。该掩码并非真实的实例分割掩码，而是以白色圆圈标记花粉团的中心，每个白色圆圈均带有黑色边框，确保白色圆圈之间无重叠。这一设计可帮助U-net学习花粉团的分割边界，简化U-net输出特征图上的团块检测或局部极大值提取流程。 val/imgs/coords/：为每张图像生成包含标注花粉团中心坐标的CSV文件。