Machine learning hackathon in Saint-Petersburg
⭐ Star us on GitHub — it helps!
YOLO network transfer learning is a good way for train the network for object detection. We implemented transfer learning for YOLO2 network to detect banners in the photoes of Saint-Petersburg.
The main idea of the task is to improve the view of the city: today the advertisement became a big problem for Saint-Petersburg, as there are lots of ugly banners in the city:
We prepared the dataset for 3 classes of banner: signboard; win_ad; cantilever. We also created description for the objects in images as YOLO needs.
Annotation:
signboard 0
win_ad 1
cantilever 2
The dataset with descriptions can be downloaded here (https://drive.google.com/drive/folders/11lwA99sJSNqYJc_d5IMEyavsHa_PxUAJ?usp=sharing).
Download Yolo_Darknet_yolo_train.ipynb and run in google collab step by step. You should download the data into the folder data/obj in google collab. Training set and test set were splitted 90% to 10%.
After 700 of epochs we achieved an impressive results:
Signboard:
Win_ad:
Cantilever:
