I had a requirement of finding a movie poster, normally found in the streets, walls or everywhere. They are usually pasted in groups of many different movie posters. My job was to locate that one poster, from the group.
I am supplied with the group photo (clicked with a camera from far). So, under different angles and different lighting conditions. This captured photo not only has the photo that I need to identify, but all other posters, pasted nearby. Also, I am supplied with the original photo, that I use for training purpose.
Initially, I have tried with several API’s from Google and Amazon (Rekognition). None of the API’s could predict or nowhere close. Than I started with ML using Tensor flow and started training. It was about 40% - 50% correct but consumed heavy GPU.
Next, I switched to OpenCV and it’s prediction is almost 95% correct. I want to know what algorithm or principle is behind OpenCV or is there a better approach.
Comments
Post a Comment