Welcome
Hi, I am Zhiying Du. This is my project report of Google Summer of Code(GSoC) 2016. My project is "Motion-based Retrieval: Feature Extraction", a subproject of vitrivr.org.
Introduction
vitrivr is an open source full-stack content-based multimedia retrieval system. it composed of Cineast(the feature extraction and retrieval engine), ADAM(the database engine) and vitrivr-ui(the UI front-end).
The aim of this project is to optimize the current motion-feature-based retrieval, including creating new motion features, changing the retrieval process and UI front-end accordingly.
Main work
- More robust path generator with forward-backward point tracking
- Foreground/background separation using RANSAC algorithm
- Foreground masking using nearest neighbor method and 3D Gaussian filter
- Foreground/background switch in the front-end
Result
- Feature extracton: Point tracking and foreground masking
| Operation | Time Usage(Single Frame) |
|---|---|
| Point tracking | 28ms |
| FG/BG separation | 6ms |
| Histogram calculation | 1ms |
| Masking | 138ms |
| Total | 201ms |
- Motion-based retrieval
Code
GSoC Project Page
Motion-based Retrieval: Feature Extraction