You're going to work on a working screw detector which somewhat works, and your task will be to improve the detection rate of the detector. Currently the detector uses hough transform to filter the candidates, and then conducts image classification (using inception v3) on the candidate ROIs to better filter the results, and as a last step, it collects 20 frames and conducts hierarchical clustering to eliminate the false positives. All in all, it's a pipeline that works to a degree.
The problem, however, is the following: Since the image classifier is trained with positive and negative images, and since these images are collected under certain lightening condition, the network cannot generalize. I'll share the dataset with you, and you can add on it as you wish. I collected the data by simply saving the candidates thrown out by the Hough transform, because this throws out screws (positive images) and screw-like artifacts (negative images). However, if you believe a better network could be employed here, let me hear what your opinion is.
This is a computer vision project so only those who have experience in the field should really consider it. I am using ROS at the moment (connecting to a camera via ROS, getting images and processing them) but you can take out the ROS part for yourself while developing, but you need to integrate it before the delivery.
So, here is what I give you:
- The code: Python and ROS melodic required. (Or if you know how to work without ROS, this can be ignored as well).
- Dataset: Screws and artifacts, small images of all.
What I expect:
- Screw Detector which works with high accuracy, under slightly different lightening conditions.
About the recuiterMember since May 20, 2018 Kumar Piyush
from al-Fayyum, Egypt