Pollen13k

The built dataset consists of more than 13 thousands per-object images collected from aerobiological samples, classified into five different categories:

• Corylus avellana, well-developed pollen grains
• Corylus avellana, anomalous pollen grains
• Alnus, well-developed pollen grains
• Debris (bubbles, dust and any non-pollen detected object)
• Cuprissaceae

Airborne pollen grains were sampled by volumetric spore traps (Lanzoni VPPS® Hirst-type sampler), which were placed at plant canopy level. The volumetric traps (air suction rates = 10 l min-1) contained adhesive tapes able to capture pollen grains and other airborne particles, that were placed on a rotating drum moved at 2 mm h-1 under a suction hole. Tapes were weekly inspected and cut into daily segments to be placed on microscope slides with a mounting medium containing basic fuchsin (0,08 % gelatin, 0,44% glycerin, 0,015% liquefied phenol, 0,0015% basic fuchsin in aqueous solution) to selectively stain pollen walls 14. Images were manually acquired from microscope slides with a Leitz Diaplan bright-field microscope and a five MP CMOS sensor. Acquired images were affected by a heavy background noise deriving either from the aerobiological sample itself (i.e. debris and dust and fungal spores) or from the mounting technique (air bubbles). For this reason, the dataset includes four pollen types, and an additional category named “Debris”. A segmentation pipeline was developed to locate and segment the grains from the microscope images. Then, experts in aerobiology manually labelled each extracted object.
The dataset includes the following files:

• An 84x84 RGB image for each segmented object, for each of the five categories;
• binary mask image for single object segmentation (84x84 resolution);
• segmented images with black or green background (84x84 resolution).

In the figures below, an example of each considered class is reported. In particular, for each class, the following images are shown (from left to right): RGB patch including the segmented object extracted from the original microscope image, binary mask defined by the segmentation procedure, segmented object with black or green background.

Figure 1 – Example of class 1 (Normal Pollen)

Figure 2 – Example of class 2 (Anomalous Pollen)

Figure 3 – Example of class 3 (Alnus)

Figure 4 – Example of class 4 (Debris)

Figure 5 – Example of class (Cuprissaceae)

Research Papers

• Battiato, S., Ortis, A., Trenta, F., Ascari, L., Politi, M., & Siniscalco, C. (2020, October). Pollen13K: A Large Scale Microscope Pollen Grain Image Dataset. In 2020 IEEE International Conference on Image Processing (ICIP) (pp. 2456-2460). IEEE.
• Battiato, S., Ortis, A., Trenta, F., Ascari, L., Politi, M., & Siniscalco, C. (2020). Detection and Classification of Pollen Grain Microscope Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (pp. 980-981).

Events

The POLLEN13K dataset has been used on the Pollen Grain Classification Challenge at ICPR 2020 challenge website

Download the dataset

The set of data used for the Pollen Grain Classification Challenge (ICPR 2020) is now available!

Researchers and educators who wish to use the Pollen13K dataset for non-commercial research and/or educational purposes must cite the above papers in their works.
Train set (MD5: 5ad8fbd0897f48bd8a68567f25f66093)
Test set (MD5: 3ca94e84059efc116bc39eda615638e2)
Test labels