timm/bee-dataset

Description

BeeDataset, from the BeeAlarmed camera based bee-hive monitoring project. This is the 300x150 (h x w) pixel image variant.

' This dataset contains images and a set of labels that expose certain characteristics of that images, such as varroa-mite infections, bees carrying pollen-packets or bee that are cooling the hive by flapping their wings. Additionally, this dataset contains images of wasps to be able to distinguish bees and wasps.

The images of the bees are taken from above and rotated. The bee is vertical and either its head or the trunk is on top. All images were taken with a green background and the distance to the bees was always the same, thus all bees have the same size.

Each image can have multiple labels assigned to it. E.g. a bee can be cooling the hive and have a varroa-mite infection at the same time. '

Labels:

• cooling - the bee is cooling the hive, flapping its wings while keeping its position stationary to move fresh air into the hive.
• pollen - the bee carries a pollen packet.
• varroa - the bee is infested with the varroa mite, a small brown circular parasite 1-2mm in diameter.
• wasps - a wasp, not a bee, was captured.

Images with none of the above (a plain, unremarkable bee) have an empty labels list.

The original dataset has no predefined splits (it is train only in TFDS). I have taken 10% of it for a validation split, using stratified sampling over the label combinations.

• Website: https://raspbee.de
• Source: https://github.com/BeeAlarmed/BeeDataset

Dataset Structure

• image - JPEG image, 300 pixel height x 150 pixel width, original bytes from the source release.
• labels - sequence of ClassLabel indices of the positive labels (0: cooling, 1: pollen, 2: varroa, 3: wasps).
• label_names - names of the positive labels.
• image_id - original image filename stem.
• cooling_output, pollen_output, varroa_output, wasps_output - the per-label fields of the TFDS output feature dict, flattened, as bool (TFDS stores them as float64 0/1).

Usage

This dataset follows the same multi-label format as timm/plant-pathology-2021 for use with timm:

python train.py --dataset hfds/timm/bee-dataset --train-split train --val-split validation \
  --task multilabel --target-key labels --num-classes 4

Citation

@misc{BeeAlarmed - A camera based bee-hive monitoring,
  title =   "Dataset for a camera based bee-hive monitoring",
  url={https://github.com/BeeAlarmed}, journal={BeeAlarmed},
  author =  "Fabian Hickert",
  year   =  "2021",
  NOTE   = "\url{https://raspbee.de/} and \url{https://github.com/BeeAlarmed/BeeAlarmed}"
}