Exploring SEIDEN: Sampling With Rare Events for Indexing in Video Database Systems
Author(s)
Potter, Cameron George
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
SEIDEN is a video database management system (VDBMS) that utilizes lightweight oracle models like YOLOv5 to derive labels and generate proxy scores for efficient video query processing. To promote optimization for rare events, we analyze the relationship between anchor count, DNN invocations, and recall for specific frame ratios and object classes. The performance of SEIDEN is evaluated based on precision, recall, and F1 score using these parameters.
The results demonstrate that SEIDEN's performance is significantly affected by the frame ratio and DNN invocation values used, with interesting patterns emerging for common and rare object classes. Increasing the DNN invocation value improves recall for both classes but at the cost of decreased precision. However, the trend is more significant for the rare class, suggesting that allocating more computational resources could enhance SEIDEN's performance more substantially for rare events than for common events.
The results yield insights into the optimization of SEIDEN's performance in object detection tasks. The findings further demonstrate the potential of SEIDEN as a fast and accurate VDBMS, even for rare events. Future research could expand on these findings by examining SEIDEN's generalizability to other datasets and comparing its performance with other methods.
Sponsor
Date
Extent
Resource Type
Text
Resource Subtype
Undergraduate Research Option Thesis