80 likes | 157 Views
Dive into TRECVID's Object and Action recognition challenge, detecting key elements in images and videos with precision. Get insights from Koen van de Sande's winning strategies and explore the power of Color Descriptors for robust visual systems.
E N D
TRECVID • Object and Action recognition competition. • Must detect these objects and actions in keyframes or video. • Goal is to accurately do so.
TRECVID • This past week • Read up on Koen van de Sande • Layout for the next 2 weeks
The Week in Review • Introduced to the code and workings of TRECVID • Added logistic regression to the existing code structure. • Brought up to speed on annotating images.
Koen van de Sande • Member of the group that won it last year. • Using his group’s executable for descriptors. • Evaluation of Color Descriptors for Object and Scene Recognition • Allows for a more robust vision system.
Next 2 Weeks • Koen van de Sande’s paper mentioned SIFT descriptors as well as bootstrapping. • Read up on both. • Possibly start to get bootstrapping implemented. • Start generating confusion matrices to get a benchmark for where we are starting from. • Annotate, annotate, annotate.
Looking Ahead • Motion • We have keyframes, how do we extrapolate motion from that or analyze actual video.
Questions • How do we decide which features? • Problem of scarcity with some images. • How do we generalize when given smaller datasets? • Many more negative examples than positive.