Bachelor's Thesis | Real-time Detection of Encoded Shapes

Thesis supervisors: Prof. Matthias Rätsch, Dr. Eike Rehder, Jonas Uhrig

Common object detection methods mainly focuses on predicting bounding boxes around objects, which provides little information about the object beyond its location and scale. In this work, a simple end-to-end trainable joint object detection and shape prediction pipeline is introduced and evaluated using the Cityscapes dataset. Existing object detection models can be easily extended through this method for shape mask prediction.