DATASET CREATION In the logo_data/ folder there is a script named create_dataset.py The purpose of this script is to take the logo images contained in the logo_data/logo_images folder (in this case only the wurth.png logo) and paste them in random positions on background images contained in the logo_data/bg_images folder, after rotating and scaling them randomly. The script creates a logo_data/dataset folder containing everything needed to train a yolov3 model (training and validation set, with images and annotations). Also, the mp4 file contained in the logo_data/test_video folder is taken and each frame is extracted and saved as test set. TRAINING In the project root folder there is the train.py script downloaded from the original github repository. Training can be started with the command python train.py --cfg cfg/yolov3-logos-big.cfg --data logo_data/dataset/logo.data --img-size 832 If you have any memory problems, you can add the option "--batch-size N" where N is by default 16. The training takes roughly 20 epochs (about 15 minutes on gpu). If running on CPU i suggest adding "--device cpu" to the command, but it takes much longer to train. Alternatively, you can use a Google Colab instance at https://colab.research.google.com/drive/18j0THvYC60jc2LRGX_XoYKWPtuyKlB-w (remember to select GPU at Runtime>Change Runtime Type before running the code) The cfg/yolov3-logos-big.cfg file is a configuration file for the model, modified for this task following the instructions on https://github.com/ultralytics/yolov3/wiki/Train-Custom-Data on step 5. The latest epoch of the model is saved in weights/last.pt and can be used to do inference. To run the model on the validation set python detect.py --source logo_data/dataset/images/val/ --output logo_data/output/ --cfg cfg/yolov3-logos-big.cfg --weights weights/last.pt --data logo_data/dataset/logo.data To run it on the test set: python detect.py --source logo_data/dataset/images/test/ --output logo_data/output/ --cfg cfg/yolov3-logos-big.cfg --weights weights/last.pt --data logo_data/dataset/logo.data --img-size 1248 The result is images within the logo_data/output/ folder. If you want the video output, you can run the logo_data/make_video.py script which takes all of the images in the logo_data/output/ folder and joins them in the logo_data/output.mp4 video.