[PYTHON] I tried using Selective search as R-CNN
Hello. The recent artificial intelligence boom never stops, I have come to see a lot of artificial intelligence related news in my daily life. Higashi Robo-kun, who is aiming to enter the University of Tokyo, recorded a "deviation value of 57.8" in the center test mock exam. I took the exam a long time ago, but I wanted you to take the center exam instead (laughs). Also, that famous [Elon Musk](https://ja.wikipedia.org/wiki/%E3%82%A4%E3%83%BC%E3%83%AD%E3%83%B3%E3% 83% BB% E3% 83% 9E% E3% 82% B9% E3% 82% AF) says, "AI is like calling the devil, especially for military use." Whether or not AI is a devil, I think it's the best opinion about military use. I want you to make good use of new technology, not for war. Well, personally, I think it will be a while before a more human-like AI is born, but what will happen?
Overview
Well, last time I tried using BING (Objectness) as R-CNN, This time, I would like to use Selective Search, which has become a hot topic recently. For R-CNN and BING, please refer to Previous article.
What is Selective Search?
It is an algorithm that selects candidate areas by grouping similar areas at the pixel level. The point is that areas with similar characteristics are combined and extracted as a single object.
procedure
This time, I used the library of Alpaca. Tried environment: Windows7 (64bit), Python2.7 Install the library with "pip install selective search".
result
It is the result of the airport image as before. (* Since it is not clear to display all the detected areas, the size of the detected objects is limited and drawn.)
You're detecting something that looks like an object.
Next is the result of the intersection image.
Well, was the image a little too complicated? Or is it a parameter issue?
Summary
It feels like it has a higher accuracy than BING. However, Selective Search has the fatal weakness of being slow. (It took 20 to 30 seconds in my environment) * This is just the impression of the library used this time. I haven't tried other libraries so I don't know. The problem of processing speed may be improved by parallel processing of GPU. I look forward to future developments. ** = Can anyone please tell me how to speed up or improve the accuracy of Selective Search = **
Recommended Posts