Evaluation of Temporal Queries in Lifelog Retrieval (Bachelor Thesis, Finished)

Every year, a huge amount of data is collected that contains information about our lives. Lifelog retrieval focuses on querying this data. Many of these queries describe whole situations that contain multiple events that are temporally dependent on each other. These queries are called temporal queries and pose new challenges to multimedia retrieval systems.

In this thesis, a pool of tasks is created that contains temporal queries. This allows us to evaluate nine temporal scoring algorithms that take into account temporal dependencies in temporal queries and two retrieval features, CLIP and text co-embedding, using the retrieval system vitrivr as an example. By analyzing the quality and scoring time of the tasks for each algorithm, it was found that temporal scoring algorithms show an improvement in the results compared to segment scoring algorithms. For vitrivr, the algorithms IDA and LNA show the best results. The two retrieval features, CLIP and text co-embedding, show similar results when using vitrivr. In conclusion, we have enabled an evaluation of temporal queries in lifelog retrieval systems.