Multiclass classification

[zhengpeng7]

Multiclass classification

[ZhengPeng7]

I think it is possible, but have no time to do it now. It's a classification first and identification second task. Currently, there is not enough work on it.

[INF800]

No worries, I can implement it if you can guide me through non-technical stuff.

I am planning to test it for this use case and see how it performs.

Until now, I've made a PoC to check if we can add numbers using CNN and it seems to work alright https://www.kaggle.com/c/ultra-mnist/discussion/313981

Now, only thing that I need to do is extract the small and big digit features in high resolution image.

So, on a very high level can tell me what should be my steps? I will start working on it

[ZhengPeng7]

Well, I think this task is very different from the person search task.
In your task, it's a normal object detection (with bbox, obtain it by cv2.findContours), multi-label classification (without bbox). But to be honest, I know only a little in multi-label classification, sorry for that.

[INF800]

We can consider this as a multiclass classification (not multilabel). It is okay Even if we are not able to get high accuracy for this problem. But the main problem that we are trying to solve is to detect bboxes of multiple objects in high resolution images - single/ multi class.

[INF800]

Well, I think this task is very different from the person search task. In your task, it's a normal object detection (with bbox, obtain it by cv2.findContours), multi-label classification (without bbox). But to be honest, I know only a little in multi-label classification, sorry for that.

You can think of it as a digit search task

[ZhengPeng7]

Sorry, INF800, I have to say I have many urgents things to do :(, so I cannot help you on this topic.