Generating and linking superpixels between images for annotation purposes
Examiner: Torsten Sattler, Dept of Electrical Engineering
A main challenge within the development of autonomous vehicles is to ensure operability in different conditions. This is commonly solved by using more labeled image data, to obtain a more general model.
Availability of annotated data is however limited and expensive across the entire field of computer vision, as such it would be beneficial to streamline the annotation process to achieve a higher throughput and lower price per annotation. This thesis presents a new approach for annotating images using superpixels in combination with a method for propagating annotations forward through sequences, giving the user suggested annotations to be corrected.
A variety of structures for a Superpixel Sampling Network (SSN) were investigated and evaluated, concluding that U-Net was the best network architecture. The U-Net SSN was then connected with a proprietary propagation technique utilizing optical flow and morphological operations. Lastly, the usefulness of the method was evaluated by implementing the full pipeline with an annotation tool and running a time study on four users. The study showed that the propagation did not contribute to a better solution, it was however shown that superpixels can reduce the annotation time with a trade off in annotation accuracy.
With a correction tool enabled the method also allowed for annotations with the same quality as the baseline, although without any reductions in time. As such we believe that the proposed method could be beneficial given further development of the annotation tool.
Student project presentation
27 May, 2020, 14:00
27 May, 2020, 14:45