Pollen Video Library for Benchmarking Detection, Classification, Tracking and Novelty Detection Tasks

Dataset description This dataset contains microscopic images and videos of pollen gathered between Feb. and Aug. 2020 in Graz, Austria. Pollen images of 16 types: ...images_16_types.zip Acer Pseudoplatanus Aesculus Carnea Alnus Anthoxanthum Betula Pendula Brassica Carpinus Corylus Dactylis Glomerata Fraxinus Pinus Nigra Platanus Populus Nigra Prunus Avium Sequoiadendron Giganteum Taxus Baccata Pollen video library ...pollen_video_library.zip Each type of pollen is in a separate folder, there may be multiple videos per type. In each pollen folder, we included images cropped from the videos by YOLO object detection algorithm trained on a subset of pollen images as described in [1]. Cropped file name structure [Video file name]_[TrackingID]_[Image index of a grain]_[Frame index in video] Example, if a grain has 5 images, the file name would be: Anthoxanthum-grass-20200530-122652_0000000_001_00001.jpg Anthoxanthum-grass-20200530-122652_0000000_002_00002.jpg ... Anthoxanthum-grass-20200530-122652_0000000_005_00005.jpg Field data over 3 days are gathered in Graz in spring 2020. ...pollen_field_data.zip Version 2: For experiments of mitigating the distribution shift of pollen identification on field data, there are 5 types selected from field data and manually labeled by the expert. The data are zipped in "the manual_labeled_field_data_5_types.zip" The "images_5_types_9010_train.zip" and "images_5_types_9010_val.zip" contain 5 types selected from library data (images_16_types.zip), and these correspond to field data. The "images_3_types_for_ablation_study.zip" contains data on 3 levels of pollen grain hydration. These data are used for the ablation study of model generalization in pollen identification. Sample code to load the data and visualize the images is in ...plot_pollen_sample.py. Download and extract the file ...images_16_types.zip in the same folder as ...plot_pollen_sample.py to run the example. Dependecies: opencv numpy matplotlib Credit [1] N. Cao, M. Meyer, L. Thiele, and O. Saukh. 2020. Automated Pollen Detection with an Affordable Technology. In Proceedings of the International Conference on Embedded Wireless Systems and Networks (EWSN). 108–119. @inproceedings{namcao2020pollen, title = {Automated Pollen Detection with an Affordable Technology}, author = {Nam Cao and Matthias Meyer and Lothar Thiele and Olga Saukh}, booktitle = {Proceedings of the International Conference on Embedded Wireless Systems and Networks (EWSN)}, pages={108–119} month = {2}, year = {2020}, }

数据集说明

本数据集包含2020年2月至8月于奥地利格拉茨采集的花粉显微图像与视频数据。

16类花粉图像的相关数据包为"images_16_types.zip"，涵盖的花粉种类包括：Acer Pseudoplatanus、Aesculus Carnea、Alnus、Anthoxanthum、Betula Pendula、Brassica、Carpinus、Corylus、Dactylis Glomerata、Fraxinus、Pinus Nigra、Platanus、Populus Nigra、Prunus Avium、Sequoiadendron Giganteum、Taxus Baccata。

花粉视频库的相关数据包为"pollen_video_library.zip"。

每一类花粉单独存放于一个独立文件夹中，单类花粉可对应多个视频文件。每个花粉文件夹中，均包含通过YOLO目标检测算法从对应视频中裁剪得到的图像；该算法基于文献[1]中所述的部分花粉图像子集训练得到。

裁剪后图像的命名规则为：[视频文件名]_[跟踪ID]_[花粉颗粒图像索引]_[视频帧索引]。

示例：若某一花粉颗粒对应5张图像，则其文件名格式如下：
Anthoxanthum-grass-20200530-122652_0000000_001_00001.jpg
Anthoxanthum-grass-20200530-122652_0000000_002_00002.jpg
……
Anthoxanthum-grass-20200530-122652_0000000_005_00005.jpg

2020年春季，研究团队于格拉茨采集了为期3天的野外花粉数据，相关数据包为"pollen_field_data.zip"。

V2版本：为开展缓解野外花粉识别任务中分布偏移问题的相关实验，我们从野外数据中筛选出5类花粉，并由领域专家进行人工标注，相关数据包为"manual_labeled_field_data_5_types.zip"。

"images_5_types_9010_train.zip"与"images_5_types_9010_val.zip"包含从图库数据（即"images_16_types.zip"）中筛选出的5类花粉，且这5类花粉与野外数据中的对应种类完全匹配。

"images_3_types_for_ablation_study.zip"包含3种不同水化程度的花粉颗粒数据，该数据集用于开展花粉识别模型泛化能力的消融实验研究。

数据加载与图像可视化的示例代码存于"plot_pollen_sample.py"文件中。若要运行示例代码，请将"images_16_types.zip"解压至与"plot_pollen_sample.py"相同的目录下。

依赖库：opencv、numpy、matplotlib

致谢
[1] Cao N, Meyer M, Thiele L, Saukh O. 2020. 基于低成本技术的自动化花粉检测 (Automated Pollen Detection with an Affordable Technology). 收录于：国际嵌入式无线系统与网络会议（EWSN）论文集，第108-119页。
对应的BibTeX引用格式如下：
@inproceedings{namcao2020pollen,
title = {基于低成本技术的自动化花粉检测 (Automated Pollen Detection with an Affordable Technology)},
author = {Nam Cao and Matthias Meyer and Lothar Thiele and Olga Saukh},
booktitle = {Proceedings of the International Conference on Embedded Wireless Systems and Networks (EWSN)},
pages={108–119},
month = {2},
year = {2020},
}