Shot Boundary Detection with Augmented Annotations

In recent years, deep learning approaches have been considered to provide state-of-the-art results in shot boundary detection. These approaches revolve around the need for large annotated datasets. The quality of the annotations is crucial to the robustness of the algorithm. Having graphical tools to verify the correct annotation of the original datasets, as well as the correct generation of synthetic datasets is a must. In this paper we propose a framework that allow the visual inspection of the datasets, incorporating the option of editing the annotations manually, as well as annotations from other algorithms, generating a set of augmented annotations. In addition, we benchmark the performance of TransNet in three scenarios, 1) using the datasets with their original annotations, 2) using automatically generated annotations, and 3) using the combination of the previous annotations, as augmented annotations. We conclude that the usage of augmented annotations significantly improves the network results.