| Περίληψη: | Object Detection is an essential computer vision task for the industry, as it can be used for counting and inspecting products in production lines. Collecting high-quality datasets for the training of object models is expensive and time-intensive, but when applying object detection to the context of production lines, CAD models of the objects to be detected are often available. Creating a Blender script that can generate photo-realistic images, including bounding box annotations of the object of interest is an alternative option. This Diploma thesis presents an approach of using the CAD model to render synthetic images for training Deep Learning models on object detection. Various experiments are conducted to figure out the ideal parameters for rendering and training, using the suggested metrics for object detection tasks. The experiments demonstrate that the concept of training models solely on synthetic data is promising since the approach is evaluated on real images. Our method is adaptable, which is a huge advantage for object detection in the dynamic environment of the industry. It can lead to considerable cost savings and increased productivity.
|
|---|
