- In this exercise, large version of YOLOv5 was the best model with mean average precision %83.3 (at 91th epoch). Colab was used in this exercise. For the best performance, I applied 100 epochs to 320x320 image size. The training process was very fast. It was completed in 1 hour 51 minutes with. It was seen that changing the image size has an impact on accuracy of the model.
- You may reach the used dataset via following link: http://ufldl.stanford.edu/housenumbers/
Note: For further details of the project you may read House-Number-Detection.pdf file.