CherryChèvre: A fine-grained dataset for goat detection in natural environments

Abstract

We introduce a new dataset for goat detection that contains 6160 annotated images captured under varying environmental conditions. The dataset is intended for developing machine learning algorithms for goat detection, with applications in precision agriculture, animal welfare, behaviour analysis, and animal husbandry. The annotations were performed by expert in computer vision, ensuring high accuracy and consistency. The dataset is publicly available and can be used as a benchmark for evaluating existing algorithms. This dataset advances research in computer vision for agriculture.

Fig. 1

Example of images featuring sheep from the COCO dataset.

Fig. 2

Example of images featuring sheep from the PASCAL VOC dataset.

Fig. 3

Incorrect annotation among different goat datasets in Roboflow.

Fig. 4

Normalized centroid distribution across the entire dataset. The presence of the sky and ground areas results in a less uniform sampling of goat positions along the vertical dimension (y).

Fig. 5

Distribution of normalized size, which includes both width and height, for the entire dataset. The results indicate that both dimensions are evenly distributed.

Fig. 6

Spatial distribution of bounding boxes in the images across the dataset. The color indicates the density of bounding boxes, with white indicating a high density, red a medium density and black indicating a density near zero. The figure provides insight into the spatial distribution of the animals within the images and highlights the areas where the animals are most frequently present.

Fig. 7

Example of annotated images for the Crosscall subset. The first line shows images near Albiez-Montrond and near Laguiole while the others was taken in INRAe Duclos. These images are very diverse, showing Creole sheep in high level of weed. These sheep are hard to distinguish from European goat.

Fig. 8

Example of annotated images for the Phantom3 subset. The first two lines feature goat captured from 22 meters above the ground, while the following two lines offer a variety of viewpoints from distant to close-up goats, as well as goats in interior environments, all captured by the flying drone.

Fig. 9

Example of annotated images for the Timelapse subset. These images are of lower quality, smaller and noisier. However, they have a large amount of individuals, and a lot of overlap between individuals. The angles of view, height of weeds, etc, are also different. This data set is therefore important, especially for those who wish to work with low resolution cameras, for wildlife conservation or theft prevention.

Fig. 10

Example of annotated images for the Tracking subset. It contains high-quality annotated images, which ensures the best detection quality, which is critical for studying animal movement patterns, habitat usage, and behavior.

Fig. 11

Example of annotated images for the External subset. It mostly showcases goats raised indoors, as seen in the first line which features goats proposed trough Mosar. The second line highlights outdoor grazing goats proposed trough Ferlus.