This course teaches action recognition in temporarily untrimmed videos. The learning outcomes include understanding action recognition video datasets, data statistics, video resolution, annotation pipelines, and challenges with evaluation videos. The course covers skills such as working with THUMOS test video set, using the Viper Toolkit for annotations, and dealing with ambiguous video segments. The teaching method involves lectures on dataset statistics, annotation pipelines, and future benchmark requirements. The intended audience for this course includes students and professionals interested in computer vision, machine learning, and video analysis.
Overview
Syllabus
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
Taught by
UCF CRCV
