Needs advice

I am trying to get started with using TensorFlow. I have a few projects that mainly focus on image classifiers in mind but I have a question on a project that I'm trying to possibly do. The problem that I am hoping to solve is to find the orientation of an object. I was wondering if I have an object such as a cylinder and the problem is to find the orientation to insert into a machine via an arm. I am working with others who are more familiar with python but I don't know if TensorFlow would have the capabilities for my project. I think it would be able to identify the object but a solution I ponder is if it would be possible to run one image classifier to find and box the cylinder and another to find the orientation since the input would be the boxed image, and then to find a rudimentary orientation based off of many pictures with them at various angles. Or if this level is too complicated to run accurately with reasonable speeds. Possibly there is a much simpler solution or is a very hard problem. Though I thought I would ask to see if there is something else that would better suit my needs before taking the time to learn TensorFlow and try to make it only to be told it won't work.

5 upvotes·13.7K views
Replies (3)

I think it might be not so important which tensor computation framework you choose. It might be a better idea to see which model(s) fit your needs best and then have a look, where you can implement them with the lowest barrier. I would suggest you have a look at YOLO (You Only Look Once) and YOLO v2 or v3 and see if it fits to your requirements. From a fast look, I would say it should do the trick:

YOLO 3 seems to have moved to Pytorch, though:

3 upvotes·7.6K views
Senior Software Engineer ·

You can use an object classification coupled with a detection model and then train it on a set of images that can identify your object. To train the classification model, you can edit and expand the training set using opencv as well.

2 upvotes·12.3K views
View all (3)
Avatar of noliemoorman2006