A Framework for Scalable Trainable Image-based Query in Video Jianbo Shi Informedia-II Carnegie Mellon University
Informedia Project
Lesson learned
Scaling Video Image Processing Engine
Human interests
PPT Slide
Body Detection/Activity Recognition in Video
The main difficulty:
Existing Approach
Body finding in a video:
Proposed approach
Joint location from motion field
Body joint tree:
Clique tree
Possible body-part locations:
Some of the possible body configurations:
Optimal configuration:
Efficient Representation of Video
Image Registration:
Multiple layers
Motion decoupling
Motion Segmentation with Normalized Cuts
Rapid training of new objects in Video
Multi-level object representation/training
Learning Segmentation
Learning Seg. With Random Walk
Learning Convexity
Learning Verbal image query
Image learning from text
Conclusion
Home Page: http://www.informedia.cs.cmu.edu