You are going to extend an existing screw detector that detects the available screw heads on devices surfaces. Your task will be to make this detector also classify the detected screw heads based on their types (slot, Torx, etc.). I'll be given you the dataset when I collect it. Essentially, what I'm going to do is to collect these different kinds of screws and take & crop images of those. So in the end, you'll be folders of images of those types. At the moment, I am not sure how many images however for the detector we needed 10.000 images (positive and negative) to train. Since this is a classifier that works on more classes, I would say we need way more than that. It also depends on how many different screw types we want to classify, which I am going to make clear towards the end of this week. But for now, we can assume this number to be 20. So, we are talking about a classifier that will classify 20 different types of screws. You can start thinking on the plan of attack at this point.
What I give:
- Training dataset. This will be consisting of cropped screw head images of thousands, taken in different lightening conditions from above.
- Screw detector. This is a trained network that finds the screws on a device with high accuracy. I have it working on ROS. If you don't know or don't have ROS, then I can give you the standalone Python code.
What I expect:
- Implementation of the classifier. You may get the output from the detector, and feed it into your pipeline.
- You are going to pick a few state-of-the-art classifiers (or you can ensemble your own model as well) and conduct evaluation on the given test set. This is required since you need to justify why you picked a certain classifier over others.
Conditions & Terms:
- We'll have 2-3 milestones, but be sure that most of the payment will be done when you successfully hand out the evaluation results. If the classifier you implemented hits low accuracies then you won't be paid. I set the threshold for the minimum payment to be 70% for all types of screws. Meaning that you classify these screws with such accuracy minimum to get paid.
- For every 10% of higher accuracy, I'll pay 100 dollars more. Say you hit 80% for all types, you'll get 100 dollars more. If you hit 90% for all types, then 200 dollars more and so on.
- I take the entire code written, including the model itself.
About the recuiterMember since May 20, 2018 Fathana Afiata
from Hunedoara, Romania