The VBS was started with a very small dataset (38 hours) and search in single video files only. However, over the years it switched to collection search and incrementally increased the dataset size to currently 1,000 hours of videos (about 1 million shots; in the V3C1 dataset – the Vimeo Creative Commons Collection – Part 1). This will be further increased to about 2,300 hours (V3C1 and V3C2, with about 2.5 million shots).
Since the first VBS in 2012, an incredible number of 30 unique teams from 21 different countries (when considering the corresponding author only) have participated in the VBS. Several teams have participated in many years, some even with different sub-teams.
VBS Winners and Teams
The following picture shows the winners of the VBS from 2012 until 2020:
Here are a few more screenshots of other great tools that participated in the VBS:
The setup of the VBS is shown in the next picture:
- All participants are sitting in a u-shape manner in front of the projector screen that shows the interface of the VBS Evaluation Server („VBS Server“).
- The VBS Server shows the content of tasks, the remaining task time, and all submissions received so far (including their status) — all participating teams are connected to the VBS Server via a local network.
- The moderator introduces the next task type and comments submissions and tasks.
- For AVS tasks judges perform judgement of submissions that are not in the ground truth.
The analysis of the performance in a YOY comparison shows that the search time for visual KIS tasks actually decreased, even though the dataset size was significantly improved – this is also true for the Novice Run (except for 2020, where it was performed in a slightly different manner).
Also, the analysis shows that even for textual KIS tasks (where no visual information is available to the teams whatsoever) the search time increased only slightly. These results clearly demonstrate that the VBS teams use highly sophisticated video retrieval tools, which allow for quick content search in large video collections.