GANs for Object Detection
This semester I had a seminar project which is about a topic of choice.
I chose to complete the seminar at an external company to get some more experience and insights as I only got experience from my training company until then. While I first aimed high and tried to dive into DeepL.com, which is a popular translation service using machine learning, I have been accepted at another machine-learning company in Aachen. The company provides auxiliary driving technology for trucks which detects if somebody is in the blind angle and warns. They also have some other projects but this was the project I had my seminar in.
The technology used to detect if a human is standing or cycling in the blind spot is based on image recognition. Other solutionas with temperature or distance based systems are also possible but have a lot of false positives.
My seminar was about a possible way to improve the object detection in uncommon situations which don’t have a lot of data. The base principle is to use the available data of the situation to train a GAN (a special neural network), which can convert pictures of common situations into uncommon situations.
With the artificially created (or better said transferred) images, a new object detector could be trained. The new object detector could now be better in detecting object in the uncommon situations, as it was trained with more images (the original ones and the additional artificially created).
In my case it worked and gave slighlty improved results, but I could not test this enough to tell if those results are statistically significant or just “luck”. The used object detection network is tinyYOLOv4. The project involved about 1000 gpu training hours on a RTX 2070. As usual when working with GANs, image transfers are working quite well when the same objects are visible. Converting a tree into a car would not work very well (converting a beetle into a car would work rather well, for historic reasons ;) ).
The project gave me experience in the use of generative adversarial networks and object detectors and helped me with writing scientific papers. It’s quite bad that I could not test significance of the results, but I am anyway more happy with the new know-how I aquired during this project.
I especially want to thank my adviser for providing me the possibility to work on this project.
Links: