Synthetic Waste Generator for Classification Training

Synthetic Waste Generator for Classification Training

El Haji, Khalid (TU Delft Electrical Engineering, Mathematics and Computer Science)
Posner, Noah (TU Delft Electrical Engineering, Mathematics and Computer Science)
Ilbaş, Hakan (TU Delft Electrical Engineering, Mathematics and Computer Science)
Karpuz, Sergen (TU Delft Electrical Engineering, Mathematics and Computer Science)
Wernet, Victor (TU Delft Electrical Engineering, Mathematics and Computer Science)

Chen, Lydia (mentor)

Delft University of Technology

2020-07-01

As the population increases so does the waste that is generated. Manually recycling waste is expensive and slow. Computer Vision (CV) solutions aim to make this less expensive and faster. Lots of data of this waste (thousands of images) is needed to train these CV solutions. This project, called Synthetic Waste Generator (SWaG) can create synthetic waste data through the use of Blender and Python. Moreover, this project makes a contribution to the current state of research by having developed an automated synthetic data generation pipeline. This synthetic data can be used to train CV solutions to enable automated recycling procedures. With the help of adjustable parameters, the synthetic data can be customized, such that different unique images of waste can be created deterministically based on a seed. Furthermore, SWaG is fully portable as it has been containerized using Docker which makes it extremely easy to obtain even faster results by running SWaG on an NVIDIA GPU enabled system as a single local container, on the cloud as a farm or incorporate it in a container-orchestration system such as Kubernetes. SWaG also crushes 3D models, to mimic real waste using soft body dynamics. The pipeline has also been suited to automatically generate COCO datasets by using masking and image segmentation techniques. SWaG can also add textures and different colors to the waste objects in the synthetically created image. Furthermore, with SWaG different conveyor belt setups at recycling plants can be simulated with the help of variable camera heights, conveyor belts, backgrounds and lighting conditions. SWaG is currently deployable and is being used and built upon by our client. After conducting empirical research experiments with SWaG, it is noted that its performance speed is linear as the amount of objects that are in a given scene increases. In fact, with between roughly 40 and 80 objects SWaG performs sub-linearly. This is an important performance criteria as images of trash on the conveyor belt often have tonnes of objects pilled up on top of one another.

Computer Vision
Recycling
Synthetic Data Generation
Generative Adversarial Networks

http://resolver.tudelft.nl/uuid:134ccb70-0139-438d-a823-837ef59087fd

Student theses

bachelor thesis

© 2020 Khalid El Haji, Noah Posner, Hakan Ilbaş, Sergen Karpuz, Victor Wernet