The COYL-1 dataset: Counting Orange and Yellow midge Larvae

1. Overview This dataset contains 369 labelled RGB images of larvae of orange wheat blossom midge, Sitodiplosis mosselana (Géhin), and yellow wheat blossom midge, Contarinia tritici (Kirby). Each image represents a plastic cup with larvae floating in water. An image could contain 3 to 1,010 larvae with an average of 182 larvae per image. Some images contained only S. mosselana larvae, some images contained only C. tritici larvae and some images contained larvae of both species. Each larva was labelled using a bounding box. Note : This dataset was used to train, validate and test a model to separately detect and count larvae of S. mosselana and C. tritici. Out of the 369 images, only 254 were retained for the train, val, and test sets. This selection was made to exclude images containing only C. tritici, the majority class. 2. Material and methods Orange and yellow larvae were imaged in different experimental settings, generating a large diversity of images. The variations between experiments concerned mainly the type of smartphone used and the type of larvae in the images (only orange larvae, only yellow larvae, or a mix of orange and yellow larvae). The common features between the experiments are listed below. In all experiments, the larvae floated in water in a transparent plastic cup. The images were taken in a greenhouse with an obstructed roof to avoid direct sunlight. The distance between the camera of the smartphone and the larvae was around 0.2 m, although small variations were generated on purpose to increase image diversity. The smartphone was held in the hand, there was no structure to stabilize it. The water cup contained larvae but sometimes also impurities such as fragments of wheat heads and aphids. Some of these elements presented shapes or colorations similar to the larvae (e.g. wheat stamens with elongated structures resembling larvae, or aphids showing orange tones), which introduced additional visual complexity. In spring 2024, 107 images were acquired using an Iphone 11 (Apple, USA) yielding .jpg images of size 3024 × 4032 pixels. Those images contained only orange larvae. On January 31, 2025, different image series were taken. The Iphone 11 was used to acquire 4 images of orange larvae, 10 images of yellow larvae, and 16 images containing both orange and yellow larvae. Images of the same cup were also taken using a Samsung S21 (Samsung, South Korea) yielding .jpg images of size 3000 × 4000 pixels. The larvae were mixed in the cup between the images taken by the two smartphones. It should be noted that these larvae had been preserved in alcohol since spring, which may have caused small changes in shape and color. In spring 2025, 201 images were acquired using a Huawei (Huwaei, China) yielding .jpg images of size 3456 × 4608 pixels. All those images contained both orange and yellow larvae. The 369 images were labeled using the online platform CVAT.ai. Considering all the images, 67,000 larvae were manually identified by drawing bounding boxes. The orange and yellow larvae were labelled as two different classes. The labelling was performed by 12 people, but a designated supervisor ensured quality checks and inspected all the labelled images. 3. Description of the files The archive contains three folders. The folder "all_labelled_images" is the main folder of the dataset. It contains the raw images and the labels. The folder "trained_model" is an additional repository containg the weights of a trained neural network, which can be used to detect larvae. The folder "scripts" contains the Python codes. 3.1. Images and labels Image names consist of three parts separated by underscores. The first part of the name is the id of the smartphone. As there were only three smartphones, of different brands, each smartphone was identified by its brand (iphone, samsung or huawei). The second part of the name tells the type of larvae in the image (orange = only orange larvae, yellow = only yellow larvae, mix = both orange and yellow larvae). The third part of the name is the image number. Example : huawei_mix_54.jpg Labels are contained in .txt files that have the same names as the corresponding images. Each .txt file has 5 columns separated by spaces, as illustrated below. The first column specifies the class of the object: larva of S. mosselana (0) or larva of C. tritici (1). The four next columns describes the coordinates of the rectangular bounding box using the YOLO format: <x_center> <y_center> <width> <height>. The coordinates are normalized from 0 to 1. Extract from a .txt file : 1 0.415911 0.501394 0.015457 0.0210920 0.704949 0.434978 0.015304 0.0228081 0.168362 0.541163 0.027859 0.0172270 0.725720 0.424000 0.021424 0.012626 3.2. Trained model The file model_weights.pt contains the trained neural network. The file model_description.txt is a short description of the model and its performance. More details can be found in the research paper. The file train_val_test_split.csv is a list of the 254 images used to train, validate and test the model. 3.3. Scripts The three main scripts are: train.py : train the model predict.py : inference = use the model to detect and count larvae in any set of images compare_models : calculate all the performance metrics Each of these main scripts has a config script that is used to change parameters before each run. The file two_classes_config.yaml is also a config, required by the Ultralytics library. It specifies the directories of the images for the train, val and test sets.