This project aims to detect mosquito breeding grounds using YOLOv7 with a Transformer Prediction Head (Swin Transformer) using pre-trained weights based on VISDRONE dataset. The dataset used consists of a subset of videos 2, 6, 9, and 12 from the Mosquito Database, which has been pre-processed and augmented.

The dataset used for training consists of 5094 images, which have been resized to from 3840x2160 to 1536x864 and/or 640x640 (two versions) and augmented with the following settings:

• Outputs per training example: 2
• Bounding Box Rotation: Between -15° and +15°
• Bounding Box Shear: ±15° Horizontal, ±15° Vertical

The pre-processed dataset is available at https://universe.roboflow.com/luis-augusto-silva-bq4bv/mosquito-suh0p/dataset/1.

This model identify 6 classes

• bottle
• bucket
• pool
• puddle
• tire
• water tanks

To run this project, you will need:

• Python 3.x
• PyTorch
• OpenCV
• YOLOv7

You can install the required packages by running:

$ git clone https://github.com/luisaugustos/YOLOv7-MBG
$ cd YOLOv7-MBG
$ pip install -r requirements.txt

To use the pre-trained model for mosquito breeding ground detection, you can run the following command:

$ python3 detect.py --weights runs/train/yolov7x_ep300_bs20_mosquito_first9/weights/best.pt --conf 0.25 --img-size 640 --source /Users/luisaugustos/Downloads/dataset_mosquito/video10.avi

train.py allows you to train new model from strach.

$ python3 train.py --name yolov7x_ep300_bs20_mosquito_first --batch 16 --workers 4 --epochs 300 --data mosquito-1/data.yaml --weights yolov7x_training.pt --cfg cfg/training/yolov7x.yaml

This project was inspired by the Mosquito Database and built on top of the YOLOv7 implementation.

Special Thanks to @elloza and your big server. Thanks to all friends from ESALAB Team too.

Thanks to their great works

• YOLOv7
• SwinTransformer
• VisDrone