Section: New Results

Evaluation of Event Recognition without Using Ground Truth

Participants : Ramiro Diaz, Carlos Fernando Crispim Junior, François Brémond.

Keywords: Computer Vision, Event Recognition, Video Summarization.

The main goal of the work is to improve the Event Recognition process and to improve the way we build the event models as well. The work concerns the Valrose Nursing Home, it consists in monitoring older people with health issues like Dementia and who are in need of care and stimulation.

Since the video dataset contains data for about 8 months, a new evaluation method is required to properly analyze the whole dataset and gain a deeper understanding of it. Our approach consists in displaying the data in a way that can be useful either for doctors, as well for engineers to detect failures and to improve the event recognition process in an efficient way. Because of this need, a new evaluation tool has been developed and named Event Plotter.

This tool provides a new method for event evaluation. First of all, as we do not have ground truth information for the total duration of the 8 months, but just for one week, so another method is needed to check the event model efficiency. To address this issue, the tool displays all the events in the desired time period (as clusters on a timeline) and single events (or time intervals) can be selected to quickly check the video and visualize the results of the event recognition working on the fly -see Figure 18 -. The goal of this work is to understand how event recognition works, change the models on the fly, import them, and see how the recognition changes in real time.

Also, to compare new event models with the old ones, video summarization is implemented as well. Event based video summarization is utilized here to check how the recognition of one particular event type changes globally on the whole video and to display the recognition results. Also video summarization can be useful for doctors to check the way patients behave, for example playing all the videos of event ”Get-up from bed”, trying to predict patterns.

The data processed to address this issue was 1 week, because it was the time corresponding to the ground truth data. With these processed data, we have tested the efficiency of the Event Plotter tool, and we are currently improving the event recognition process by changing event models, adding new zones, and testing them on the fly.