Action Recognition in Temporarily Untrimmed Videos

Learn about action recognition in temporally untrimmed videos through this 26-minute lecture from the University of Central Florida. Explore the THUMOS 2014 dataset, its collection process, and annotation pipeline. Discover challenges in video data evaluation, including spatial and temporal resolution considerations. Examine the use of tools like the Viper Toolkit for annotation and understand the complexities of labeling positive and background examples. Gain insights into handling ambiguous video segments and the intricacies of searching for disambiguated content. Delve into the challenges of using Freebase topics and the specific requirements for future benchmarks in action recognition.

Intro
THUMOS 2014
Outline
Action Recognition in Temporally Untrimmed Videos
Action Recognition Video Datasets
What about TRECVID?
Statistics of the Data Partitions
Video Resolution (Spatial & Temporal)
THUMOS Test Video Set
Validation Data Collection
Annotation Pipeline: Positive Examples
Viper Toolkit
Annotation Pipeline Background Examples
Example Positive Annotations
Ambiguous Video Segments Examples (Excluded from Eval)
Searching for Disambiguated Things
Freebase Topics Gotchas
Challenges with Evaluation Videos (1)
THUMOS Benchmark Choices
Future Benchmark Requirements